Files
zeekay 1f308322ab Keep wide tables and specs inside a phone screen
At 390 pixels the blog's tables, the datasets tables and the model
card specs ran past the right edge, and body's overflow-x clip cut
them off rather than letting them scroll. Tables now scroll in their
own box, and a spec column shrinks and wraps instead of pushing its
neighbour out of the card.
2026-09-30 02:41:46 -07:00

1 line
117 KiB
HTML
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
<!DOCTYPE html><!--_BoRG_r6Z3v9udL0lI0J_--><html lang="en"><head><meta charSet="utf-8"/><meta name="viewport" content="width=device-width, initial-scale=1"/><link rel="stylesheet" href="/_next/static/css/1d1f6bc532e5f43f.css" data-precedence="next"/><link rel="preload" as="script" fetchPriority="low" href="/_next/static/chunks/webpack-f43ad49c79db3299.js"/><script src="/_next/static/chunks/4bd1b696-c023c6e3521b1417.js" async=""></script><script src="/_next/static/chunks/255-566af14aff71e752.js" async=""></script><script src="/_next/static/chunks/main-app-7eaab9f523b3767f.js" async=""></script><script src="/_next/static/chunks/619-ba102abea3e3d0e4.js" async=""></script><script src="/_next/static/chunks/app/layout-425ca81883167c16.js" async=""></script><script src="/_next/static/chunks/app/blog/page-f19d17b0f66e456e.js" async=""></script><script src="/assets/js/main.js" async=""></script><link rel="icon" type="image/svg+xml" href="/favicon.svg"/><link rel="alternate icon" href="/favicon.png"/><title>Zen Blog — research notes &amp; releases</title><meta name="description" content="Architecture deep-dives, training research, and release notes from Zen LM, the open model family of Zoo Labs Foundation."/><meta name="keywords" content="AI, LLM, Agentic AI, Code Generation, Zen Coder, Multimodal, Open Source, Machine Learning"/><script src="/_next/static/chunks/polyfills-42372ed130431b0a.js" noModule=""></script></head><body><div hidden=""><!--$--><!--/$--></div><header><nav class="navbar"><div class="container"><div class="logo"><a href="/"><svg class="enso logo-img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 100 100" width="34" height="34" fill="none" role="img" aria-label="Zen ensō"><defs><linearGradient id="_R_adb_" x1="12%" y1="0%" x2="88%" y2="100%"><stop offset="0%" stop-color="#ffffff"></stop><stop offset="50%" stop-color="#d4d4d8"></stop><stop offset="100%" stop-color="#7c7c82"></stop></linearGradient></defs><path class="enso-stroke" pathLength="1" d="M58 13 C38 9 17 22 13 44 C9 68 26 89 50 90 C74 91 91 74 89 50 C88 41 87 35 84 30" stroke="url(#_R_adb_)" stroke-width="7" stroke-linecap="round" stroke-linejoin="round"></path></svg><div class="logo-text"><h1>Zen LM</h1></div></a></div><div class="header-right"><div class="nav-links"><a class="" href="/">Home</a><div class="nav-item"><a class="" href="/models/">Models<svg class="nav-caret" width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M6 9l6 6 6-6"></path></svg></a><div class="nav-dropdown"><div class="nav-dropdown-inner"><a class="dd-item" href="/models/#zen6"><div class="dd-item-text"><span class="dd-item-label">Zen 6<span class="dd-tag">Available</span></span><span class="dd-item-desc">27B dense — text, images and video, 1M context with YaRN</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="/models/#zen6"><div class="dd-item-text"><span class="dd-item-label">Zen 6 Flash<span class="dd-tag">Available</span></span><span class="dd-item-desc">The ternary Zen 6 — 5.95 GB, fits an Apple Silicon laptop</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://hanzo.ai/research-access" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Zen 7<span class="dd-tag">Research preview</span></span><span class="dd-item-desc">No weights yet — request access</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="/models/#zen5"><div class="dd-item-text"><span class="dd-item-label">Earlier generations</span><span class="dd-item-desc">Zen 5, Zen 4 and Zen 3</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="/models/#catalog"><div class="dd-item-text"><span class="dd-item-label">All models</span><span class="dd-item-desc">Browse the full catalog</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a></div></div></div><div class="nav-item"><a class="" href="/datasets/">Datasets<svg class="nav-caret" width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M6 9l6 6 6-6"></path></svg></a><div class="nav-dropdown"><div class="nav-dropdown-inner"><a class="dd-item" href="/datasets/"><div class="dd-item-text"><span class="dd-item-label">Zen Agentic Dataset<span class="dd-tag">New</span></span><span class="dd-item-desc">10B+ tokens of real tool use and reasoning</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="/datasets/"><div class="dd-item-text"><span class="dd-item-label">Training Data</span><span class="dd-item-desc">Curation, sources, and methodology</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://huggingface.co/datasets/hanzoai/zen-agentic-dataset" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">On HuggingFace</span><span class="dd-item-desc">Download and license the datasets</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a></div></div></div><div class="nav-item"><a class="" href="/research/">Research<svg class="nav-caret" width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M6 9l6 6 6-6"></path></svg></a><div class="nav-dropdown"><div class="nav-dropdown-inner"><a class="dd-item" href="/research/"><div class="dd-item-text"><span class="dd-item-label">Research Papers</span><span class="dd-item-desc">Technical reports across the Zen family</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://papers.zenlm.org" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Papers Archive</span><span class="dd-item-desc">Full paper archive — papers.zenlm.org</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="/blog/"><div class="dd-item-text"><span class="dd-item-label">Blog</span><span class="dd-item-desc">Research notes, deep dives, and releases</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a></div></div></div><a class="active" href="/blog/">Blog</a></div><div class="logo-links"><a href="https://github.com/zenlm" target="_blank" rel="noopener noreferrer" class="icon-link" title="GitHub"><svg width="24" height="24" viewBox="0 0 24 24" fill="currentColor"><path d="M12 0c-6.626 0-12 5.373-12 12 0 5.302 3.438 9.8 8.207 11.387.599.111.793-.261.793-.577v-2.234c-3.338.726-4.033-1.416-4.033-1.416-.546-1.387-1.333-1.756-1.333-1.756-1.089-.745.083-.729.083-.729 1.205.084 1.839 1.237 1.839 1.237 1.07 1.834 2.807 1.304 3.492.997.107-.775.418-1.305.762-1.604-2.665-.305-5.467-1.334-5.467-5.931 0-1.311.469-2.381 1.236-3.221-.124-.303-.535-1.524.117-3.176 0 0 1.008-.322 3.301 1.23.957-.266 1.983-.399 3.003-.404 1.02.005 2.047.138 3.006.404 2.291-1.552 3.297-1.23 3.297-1.23.653 1.653.242 2.874.118 3.176.77.84 1.235 1.911 1.235 3.221 0 4.609-2.807 5.624-5.479 5.921.43.372.823 1.102.823 2.222v3.293c0 .319.192.694.801.576 4.765-1.589 8.199-6.086 8.199-11.386 0-6.627-5.373-12-12-12z"></path></svg></a><a href="https://huggingface.co/zenlm" target="_blank" rel="noopener noreferrer" class="icon-link" title="HuggingFace"><svg width="28" height="28" viewBox="0 0 32 32" fill="currentColor" aria-hidden="true"><circle cx="16" cy="16" r="13" fill="#d4d4d8"></circle><circle cx="11" cy="14" r="2" fill="#18181b"></circle><circle cx="21" cy="14" r="2" fill="#18181b"></circle><path d="M11 19 Q16 24 21 19" stroke="#18181b" stroke-width="1.6" fill="none" stroke-linecap="round"></path></svg></a></div><div class="nav-item try-item"><a href="https://hanzo.chat/?model=zen6" target="_blank" rel="noopener noreferrer" class="btn-try">Try Zen<svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><div class="nav-dropdown nav-dropdown-right"><div class="nav-dropdown-inner"><a class="dd-item" href="https://hanzo.chat/?model=zen6" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Zen 6<span class="dd-tag">Available</span></span><span class="dd-item-desc">Chat with Zen 6</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://hanzo.chat/?model=zen6-flash" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Zen 6 Flash<span class="dd-tag">Available</span></span><span class="dd-item-desc">Chat with the ternary Zen 6</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://api.hanzo.ai" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Zen API</span><span class="dd-item-desc">OpenAI- &amp; Anthropic-compatible. One key.</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a><a class="dd-item" href="https://hanzo.chat" target="_blank" rel="noopener noreferrer"><div class="dd-item-text"><span class="dd-item-label">Zen Chat</span><span class="dd-item-desc">Open the full chat experience</span></div><svg class="dd-arrow" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"></path></svg></a></div></div></div></div></div></nav></header><main><section class="hero"><div class="container"><h1 class="hero-title">Blog</h1><p class="hero-subtitle">Research notes, deep dives, and releases</p><p class="hero-description">71<!-- --> posts on Zen architecture, training, and the open-model ecosystem.</p></div></section><section class="catalog-section"><div class="container"><div class="blog-list"><a class="blog-card" href="/blog/bitdelta-behavioral-compression/"><div class="blog-card-meta">Feb 27, 2026<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family</h2><p class="blog-card-desc">How BitDelta (arXiv:2402.10193) compresses fine-tuned behavioral deltas to 1-bit precision, enabling the full Zen model family — nano through ultra — to share a single GPU cluster.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Quantization</span><span class="blog-tag">BitDelta</span></div></a><a class="blog-card" href="/blog/continual-learning-sure-opcm/"><div class="blog-card-meta">Feb 27, 2026<!-- --> <!-- -->·<!-- --> <!-- -->8<!-- --> min read</div><h2 class="blog-card-title">SuRe + OPCM: Production-Grade Continual Learning for Open Models</h2><p class="blog-card-desc">Deep dive on Surprise-Driven Prioritized Replay (SuRe) and Orthogonal Projection Continual Merging (OPCM) — the two SOTA techniques we use for catastrophic-forgetting-free LLM adaptation in the Zen model family.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Continual Learning</span><span class="blog-tag">SuRe</span></div></a><a class="blog-card" href="/blog/drop-upcycling-zen-mode/"><div class="blog-card-meta">Feb 27, 2026<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Drop-Upcycling and the Birth of Zen MoDE Architecture</h2><p class="blog-card-desc">How Drop-Upcycling (arXiv:2502.19261) transforms dense checkpoints into MoE models at 1/4 training cost, and how it shapes Zen MoDE — our Mixture of Distilled Experts architecture.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">MoE</span><span class="blog-tag">Architecture</span></div></a><a class="blog-card" href="/blog/gt-qlora-moe-abliteration/"><div class="blog-card-meta">Feb 27, 2026<!-- --> <!-- -->·<!-- --> <!-- -->8<!-- --> min read</div><h2 class="blog-card-title">GT-QLoRA: Uncensoring Trillion-Parameter MoE Models</h2><p class="blog-card-desc">Why standard abliteration techniques fail on Mixture-of-Experts models, and how Gate-Targeted QLoRA solves the expert routing problem at 1 trillion parameters.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Abliteration</span><span class="blog-tag">GT-QLoRA</span></div></a><a class="blog-card" href="/blog/zen4-ultra/"><div class="blog-card-meta">Jan 19, 2026<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Zen4 Ultra: 480B Parameters, 1M Token Context</h2><p class="blog-card-desc">Zen4 Ultra is our most capable model: 480B total parameters, 35B active per token, 1M token context window. Benchmark results and use cases.</p><div class="blog-card-tags"><span class="blog-tag">Models</span><span class="blog-tag">Zen4</span><span class="blog-tag">Flagship</span></div></a><a class="blog-card" href="/blog/zen-mode-architecture/"><div class="blog-card-meta">Jan 15, 2026<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Zen MoDE: Mixture of Distilled Experts</h2><p class="blog-card-desc">A technical deep dive into Zen MoDE — Mixture of Distilled Experts — the architecture underlying all Zen models.</p><div class="blog-card-tags"><span class="blog-tag">Architecture</span><span class="blog-tag">Research</span><span class="blog-tag">MoE</span></div></a><a class="blog-card" href="/blog/zen-lm-launch/"><div class="blog-card-meta">Jan 14, 2026<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Introducing Zen LM: Open Frontier Models from Hanzo AI and Zoo Labs</h2><p class="blog-card-desc">Announcing the Zen model family: 94+ open models built on Zen MoDE architecture, co-developed by Hanzo AI and Zoo Labs Foundation.</p><div class="blog-card-tags"><span class="blog-tag">Announcement</span><span class="blog-tag">Models</span><span class="blog-tag">Zen</span></div></a><a class="blog-card" data-upstream="" href="/blog/qwen3guard/"><div class="blog-card-meta">Sep 22, 2025<!-- --> <!-- -->·<!-- --> <!-- -->8<!-- --> min read</div><h2 class="blog-card-title">Qwen3Guard: Real-time Safety for Your Token Stream</h2><p class="blog-card-desc">We are excited to introduce Qwen3Guard, the first safety guardrail model in the Qwen family. Built upon the powerful Qwen3 foundation models and fine-tuned specifically for safety classificatoin, Qwen3Guard ensures responsible AI interactions by delivering precise safety detection for both prompts a</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-image-edit/"><div class="blog-card-meta">Aug 18, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Qwen-Image-Edit: Image Editing with Higher Quality and Efficiency</h2><p class="blog-card-desc">We are excited to introduce Qwen-Image-Edit, the image editing version of Qwen-Image. Built upon our 20B Qwen-Image model, Qwen-Image-Edit successfully extends Qwen-Image&#x27;s unique text rendering capabilities to image editing tasks, enabling precise text editing. Furthermore, Qwen-Image-Edit simultan</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-image/"><div class="blog-card-meta">Aug 3, 2025<!-- --> <!-- -->·<!-- --> <!-- -->12<!-- --> min read</div><h2 class="blog-card-title">Qwen-Image: Crafting with Native Text Rendering</h2><p class="blog-card-desc">We are thrilled to release **Qwen-Image**, a 20B MMDiT image foundation model that achieves significant advances in complex text rendering and precise image editing. To try the latest model, feel free to visit [Qwen Chat](https://chat.qwenlm.ai) and choose “Image Generation”.</p></a><a class="blog-card" data-upstream="" href="/blog/gspo/"><div class="blog-card-meta">Jul 26, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">GSPO: Towards Scalable Reinforcement Learning for Language Models</h2><p class="blog-card-desc">Reinforcement Learning (RL) has emerged as a pivotal paradigm for scaling language models and enhancing their deep reasoning and problem-solving capabilities. To scale RL, the foremost prerequisite is maintaining stable and robust training dynamics. However, we observe that existing RL algorithms (s</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-mt/"><div class="blog-card-meta">Jul 23, 2025<!-- --> <!-- -->·<!-- --> <!-- -->9<!-- --> min read</div><h2 class="blog-card-title">Qwen-MT: Where Speed Meets Smart Translation</h2><p class="blog-card-desc">Here we introduce the latest update of Qwen-MT (qwen-mt-turbo) via [Qwen API](https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?type=model&amp;url=https://www.alibabacloud.com/help/en/doc-detail/2840914_2.html&amp;renderType=component&amp;modelId=qwen-mt-turbo). This update builds upon the powerful Qw</p></a><a class="blog-card" data-upstream="" href="/blog/qwen3-coder/"><div class="blog-card-meta">Jul 21, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Qwen3-Coder: Agentic Coding in the World</h2><p class="blog-card-desc">Today, we&#x27;re announcing Qwen3-Coder, our most agentic code model to date. Qwen3-Coder is available in multiple sizes, but we&#x27;re excited to introduce its most powerful variant first: Qwen3-Coder-480B-A35B-Instruct — a 480B-parameter Mixture-of-Experts model with 35B active parameters which supports t</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-tts/"><div class="blog-card-meta">Jun 26, 2025<!-- --> <!-- -->·<!-- --> <!-- -->6<!-- --> min read</div><h2 class="blog-card-title">Time to Speak Some Dialects, Qwen-TTS!</h2><p class="blog-card-desc">Here we introduce the latest update of **Qwen-TTS** (`qwen-tts-latest` or `qwen-tts-2025-05-22`) through [Qwen API](https://help.aliyun.com/zh/model-studio/qwen-tts) . Trained on a large-scale dataset encompassing over millions of hours of speech, Qwen-TTS achieves human-level naturalness and expres</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-vlo/"><div class="blog-card-meta">Jun 25, 2025<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Qwen VLo: From &quot;Understanding&quot; the World to &quot;Depicting&quot; It</h2><p class="blog-card-desc">The evolution of multimodal large models is continually pushing the boundaries of what we believe technology can achieve. From the initial QwenVL to the latest Qwen2.5 VL, we have made progress in enhancing the model&#x27;s ability to understand image content. Today, we are excited to introduce a new model,</p></a><a class="blog-card" data-upstream="" href="/blog/qwen3-embedding/"><div class="blog-card-meta">Jun 4, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models</h2><p class="blog-card-desc">{/* {{&lt; video src=&quot;https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/qwen2-main-video.m4v&quot; width=&quot;100%&quot; alt=&quot;Qwen2.5 Main Video&quot; autoplay=true loop=true controls=false muted=true playsinline=true &gt;}} */}</p></a><a class="blog-card" data-upstream="" href="/blog/qwen3/"><div class="blog-card-meta">Apr 28, 2025<!-- --> <!-- -->·<!-- --> <!-- -->11<!-- --> min read</div><h2 class="blog-card-title">Qwen3: Think Deeper, Act Faster</h2><p class="blog-card-desc">Today, we are excited to announce the release of **Qwen3**, the latest addition to the Qwen family of large language models. Our flagship model, **Qwen3-235B-A22B**, achieves competitive results in benchmark evaluations of coding, math, general capabilities, etc., when compared to other top-tier mod</p></a><a class="blog-card" data-upstream="" href="/blog/qvq-max-preview/"><div class="blog-card-meta">Mar 27, 2025<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">QVQ-Max: Think with Evidence</h2><p class="blog-card-desc">Last December, we launched QVQ-72B-Preview as an exploratory model, but it had many issues. Today, we are officially releasing the first version of QVQ-Max, our visual reasoning model. This model can not only &quot;understand&quot; the content in images and videos but also analyze and reason with this informa</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-omni"><div class="blog-card-meta">Mar 26, 2025<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5 Omni: See, Hear, Talk, Write, Do It All!</h2><p class="blog-card-desc">We release **Qwen2.5-Omni**, the new flagship end-to-end multimodal model in the Qwen series. Designed for comprehensive multimodal perception, it seamlessly processes diverse inputs including text, images, audio, and video, while delivering real-time streaming responses through both text generation and</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-vl-32b"><div class="blog-card-meta">Mar 23, 2025<!-- --> <!-- -->·<!-- --> <!-- -->2<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-VL-32B: Smarter and Lighter</h2><p class="blog-card-desc">At the end of January this year, we launched the Qwen2.5-VL series of models, which received widespread attention and positive feedback from the community. Building on the Qwen2.5-VL series, we continued to optimize the model using reinforcement learning and open-sourced the new VL model with the beloved 32</p></a><a class="blog-card" data-upstream="" href="/blog/qwq-32b/"><div class="blog-card-meta">Mar 5, 2025<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">QwQ-32B: Embracing the Power of Reinforcement Learning</h2><p class="blog-card-desc">Scaling Reinforcement Learning (RL) has the potential to enhance model performance beyond conventional pretraining and post-training methods. Recent studies have demonstrated that RL can significantly improve the reasoning capabilities of models. For instance, DeepSeek R1 has achieved state-of-the-a</p></a><a class="blog-card" data-upstream="" href="/blog/qwq-max-preview/"><div class="blog-card-meta">Feb 24, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">&lt;think&gt;...&lt;/think&gt; QwQ-Max-Preview</h2><p class="blog-card-desc">This is a blog created by QwQ-Max-Preview. We hope you enjoy it!</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-max"><div class="blog-card-meta">Jan 27, 2025<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model</h2><p class="blog-card-desc">It is widely recognized that continuously scaling both data size and model size can lead to significant improvements in model intelligence. However, the research and industry community has limited experience in effectively scaling extremely large models, whether they are dense or Mixture-of-Expert (</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-1m"><div class="blog-card-meta">Jan 26, 2025<!-- --> <!-- -->·<!-- --> <!-- -->9<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens</h2><p class="blog-card-desc">Two months after upgrading [Qwen2.5-Turbo](../qwen2.5-turbo) to support context length up to one million tokens, we are back with the open-source Qwen2.5-1M models and the corresponding inference framework support. Here&#x27;s what you can expect from this release:</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-vl"><div class="blog-card-meta">Jan 25, 2025<!-- --> <!-- -->·<!-- --> <!-- -->6<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL!</h2><p class="blog-card-desc">We release **Qwen2.5-VL**, the new flagship vision-language model of Qwen and also a significant leap from the previous Qwen2-VL. To try the latest model, feel free to visit [Qwen Chat](https://chat.qwenlm.ai) and choose Qwen2.5-VL-72B-Instruct. Also, we open both base and instruct models in 3 sizes, includin</p></a><a class="blog-card" data-upstream="" href="/blog/global-load-balance/"><div class="blog-card-meta">Jan 20, 2025<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Global-batch load balance almost free lunch to improve your MoE LLM training</h2><p class="blog-card-desc">The Mixture-of-Experts (MoEs) architecture has become a popular model-parameter-scale-up technique. Typically, one MoE layer consists of a router (often parameterized as one single Linear layer) and a group of experts (for transformer-based models, each expert is one feedforward layer). Given an inp</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-math-prm"><div class="blog-card-meta">Jan 13, 2025<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Towards Effective Process Supervision in Mathematical Reasoning</h2><p class="blog-card-desc">In recent years, Large Language Models (LLMs) have made remarkable advances in mathematical reasoning, yet they can make mistakes, such as miscalculations or logical errors, leading to wrong conclusions. Moreover, even when achieving correct final answers, these powerful models can still regularly m</p></a><a class="blog-card" href="/blog/zen-3/"><div class="blog-card-meta">Jan 12, 2025<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Zen 3.0: Our Third-Generation Open Model Family</h2><p class="blog-card-desc">Announcing Zen 3.0, our most capable open model family yet.</p><div class="blog-card-tags"><span class="blog-tag">Announcement</span><span class="blog-tag">Models</span></div></a><a class="blog-card" data-upstream="" href="/blog/qvq-72b-preview/"><div class="blog-card-meta">Dec 24, 2024<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">QVQ: To See the World with Wisdom</h2><p class="blog-card-desc">Language and vision intertwine in the human mind, shaping how we perceive and understand the world around us. Our ability to reason is deeply rooted in both linguistic thought and visual memory - but what happens when we extend these capabilities to AI? Today&#x27;s large language models have demonstrate</p></a><a class="blog-card" data-upstream="" href="/blog/qwq-32b-preview/"><div class="blog-card-meta">Nov 27, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">QwQ: Reflect Deeply on the Boundaries of the Unknown</h2><p class="blog-card-desc">&lt;i style=&quot;color: grey;&quot;&gt;Note: This is the pronunciation of QwQ: /kwju:/ , similar to the word &quot;quill&quot;.&lt;/i&gt;</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-turbo"><div class="blog-card-meta">Nov 14, 2024<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Extending the Context Length to 1M Tokens!</h2><p class="blog-card-desc">After the release of Qwen2.5, we heard the community&#x27;s demand for processing longer contexts. In recent months, we have made many optimizations for the model capabilities and inference performance of extremely long context. Today, we are proud to introduce the new Qwen2.5-Turbo version, which features:</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-coder-family"><div class="blog-card-meta">Nov 11, 2024<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-Coder Series: Powerful, Diverse, Practical.</h2><p class="blog-card-desc">{/* {{&lt; figure src=&quot;https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/Qwen2.5-Coder/coder-main.png#center&quot; width=&quot;70%&quot;&gt;}} */}</p></a><a class="blog-card" href="/blog/future-of-open-ai/"><div class="blog-card-meta">Nov 10, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">The Future of Open AI</h2><p class="blog-card-desc">Reflections on where open AI development is heading and what it will take to get there.</p><div class="blog-card-tags"><span class="blog-tag">Vision</span><span class="blog-tag">Open Source</span></div></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5"><div class="blog-card-meta">Sep 18, 2024<!-- --> <!-- -->·<!-- --> <!-- -->11<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5: A Party of Foundation Models!</h2><p class="blog-card-desc">{/* {{&lt; video src=&quot;https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/qwen2-main-video.m4v&quot; width=&quot;100%&quot; alt=&quot;Qwen2.5 Main Video&quot; autoplay=true loop=true controls=false muted=true playsinline=true &gt;}} */}</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-coder"><div class="blog-card-meta">Sep 18, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-Coder: Code More, Learn More!</h2><p class="blog-card-desc">In early April, we introduced CodeQwen1.5, which garnered significant attention from the community. Since then, we have been working to enhance the coding model. Today, we are excited to announce the release of the next generation of open-source coding models, **Qwen2.5-Coder**, and officially rename Co</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-llm"><div class="blog-card-meta">Sep 18, 2024<!-- --> <!-- -->·<!-- --> <!-- -->23<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-LLM: Extending the boundary of LLMs</h2><p class="blog-card-desc">In this blog, we delve into the details of our latest Qwen2.5 series language models. We have developed a range of decoder-only dense models, with seven of them open-sourced, spanning from 0.5B to 72B parameters. Our research indicates a significant interest among users in models within the 10-30B range</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2.5-math"><div class="blog-card-meta">Sep 18, 2024<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Qwen2.5-Math: The world&#x27;s leading open-sourced mathematical LLMs</h2><p class="blog-card-desc">&gt; &lt;div align=&quot;center&quot;&gt; &gt; &lt;b&gt; &gt; 🚨 Qwen2.5-Math mainly supports solving English and Chinese math problems through CoT and TIR. We do not recommend using this series of models for other tasks. &gt; &lt;/b&gt; &gt; &lt;/div&gt;</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2-vl/"><div class="blog-card-meta">Aug 28, 2024<!-- --> <!-- -->·<!-- --> <!-- -->8<!-- --> min read</div><h2 class="blog-card-title">Qwen2-VL: To See the World More Clearly</h2><p class="blog-card-desc">After a year&#x27;s relentless efforts, today we are thrilled to release **Qwen2-VL**! Qwen2-VL is the latest version of the vision language models based on **Qwen2** in the Qwen model familities. Compared with Qwen-VL, Qwen2-VL has the capabilities of:</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2-audio/"><div class="blog-card-meta">Aug 8, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Qwen2-Audio: Chat with Your Voice!</h2><p class="blog-card-desc">To achieve the objective of building an AGI system, the model should be capable of understanding information from different modalities. Thanks to the rapid development of large language models, LLMs are now capable of understanding language and reasoning. Previously we have taken a step forward to e</p></a><a class="blog-card" data-upstream="" href="/blog/qwen2-math/"><div class="blog-card-meta">Aug 7, 2024<!-- --> <!-- -->·<!-- --> <!-- -->30<!-- --> min read</div><h2 class="blog-card-title">Introducing Qwen2-Math</h2><p class="blog-card-desc">&gt; &lt;div align=&quot;center&quot;&gt; &gt; &lt;b&gt; &gt; 🚨 This model mainly supports English. We will release bilingual (English and Chinese) math models soon. &gt; &lt;/b&gt; &gt; &lt;/div&gt;</p></a><a class="blog-card" href="/blog/decentralized-compute/"><div class="blog-card-meta">Aug 4, 2024<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Decentralized Compute for AI Training</h2><p class="blog-card-desc">How we&#x27;re building a decentralized compute network for training large AI models.</p><div class="blog-card-tags"><span class="blog-tag">Infrastructure</span><span class="blog-tag">Training</span><span class="blog-tag">Decentralization</span></div></a><a class="blog-card" data-upstream="" href="/blog/qwen2/"><div class="blog-card-meta">Jun 6, 2024<!-- --> <!-- -->·<!-- --> <!-- -->25<!-- --> min read</div><h2 class="blog-card-title">Hello Qwen2</h2><p class="blog-card-desc">After months of efforts, we are pleased to announce the evolution from Qwen1.5 to Qwen2. This time, we bring to you:</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-agent-2405/"><div class="blog-card-meta">Jun 5, 2024<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Generalizing an LLM from 8k to 1M Context using Qwen-Agent</h2><p class="blog-card-desc">**TLDR:** We&#x27;ve created an agent using Qwen2 models with an 8k context size to understand documents with 1M tokens, surpassing RAG and native long-context models. This agent was also used to generate data for training new long-context Qwen models.</p></a><a class="blog-card" href="/blog/zips-governance/"><div class="blog-card-meta">May 19, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">ZIPs: Decentralized Governance for Open AI</h2><p class="blog-card-desc">How Zoo Improvement Proposals enable community-driven governance of open AI development.</p><div class="blog-card-tags"><span class="blog-tag">Governance</span><span class="blog-tag">ZIPs</span></div></a><a class="blog-card" data-upstream="" href="/blog/qwen-max-0428/"><div class="blog-card-meta">May 10, 2024<!-- --> <!-- -->·<!-- --> <!-- -->1<!-- --> min read</div><h2 class="blog-card-title">Notes on Qwen-Max-0428</h2><p class="blog-card-desc">Previously, we opensourced a series of Qwen1.5 model ranging from 0.5 to 110 billion parameters. Now, we release a larger model, Qwen-Max-0428. Qwen-Max-0428 is an instruction-tuned model for chat service. Very recently, it is available via [Chatbot Arena](https://chat.lmsys.org/) and it has now bec</p></a><a class="blog-card" data-upstream="" href="/blog/qwen1.5-110b"><div class="blog-card-meta">Apr 24, 2024<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Qwen1.5-110B: The First 100B+ Model of the Qwen1.5 Series</h2><p class="blog-card-desc">Recently we have witnessed a burst of large-scale models with over 100 billion parameters in the opensource community. These models have demonstrated remarkable performance in both benchmark evaluation and chatbot arena. Today, we release the first 100B+ model of the Qwen1.5 series, Qwen1.5-110B, wh</p></a><a class="blog-card" data-upstream="" href="/blog/codeqwen1.5"><div class="blog-card-meta">Apr 15, 2024<!-- --> <!-- -->·<!-- --> <!-- -->7<!-- --> min read</div><h2 class="blog-card-title">Code with CodeQwen1.5</h2><p class="blog-card-desc">The advent of advanced programming tools, which harnesses the power of large language models (LLMs), has significantly enhanced programmer productivity and accuracy. Notwithstanding these advancements, dominant coding assistants like Github Copilot, built upon proprietary LLMs, pose notable challeng</p></a><a class="blog-card" data-upstream="" href="/blog/qwen1.5-32b"><div class="blog-card-meta">Apr 1, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Qwen1.5-32B: Fitting the Capstone of the Qwen1.5 Language Model Series</h2><p class="blog-card-desc">The open-source community has long sought a model that strikes an ideal balance between performance, efficiency, and memory footprint. Despite the emergence of cutting-edge models like Qwen1.5-72B and DBRX, the models have faced persistent challenges such as large memory consumption, slow inference</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-moe/"><div class="blog-card-meta">Mar 27, 2024<!-- --> <!-- -->·<!-- --> <!-- -->8<!-- --> min read</div><h2 class="blog-card-title">Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters</h2><p class="blog-card-desc">Since the surge in interest sparked by Mixtral, research on mixture-of-expert (MoE) models has gained significant momentum. Both researchers and practitioners are keenly interested in understanding how to effectively train such models and assessing their efficiency and effectiveness. Today, we intro</p></a><a class="blog-card" href="/blog/zoo-foundation-launch/"><div class="blog-card-meta">Feb 11, 2024<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Announcing the Zoo Labs Foundation</h2><p class="blog-card-desc">Today we formally launch Zoo Labs Foundation, an open research network dedicated to decentralized AI and decentralized science.</p><div class="blog-card-tags"><span class="blog-tag">Announcement</span><span class="blog-tag">Foundation</span></div></a><a class="blog-card" data-upstream="" href="/blog/qwen1.5"><div class="blog-card-meta">Feb 3, 2024<!-- --> <!-- -->·<!-- --> <!-- -->18<!-- --> min read</div><h2 class="blog-card-title">Introducing Qwen1.5</h2><p class="blog-card-desc">In recent months, our focus has been on developing a &quot;good&quot; model while optimizing the developer experience. As we progress towards **Qwen1.5**, the next iteration in our Qwen series, this update arrives just before the Chinese New Year.</p></a><a class="blog-card" data-upstream="" href="/blog/qwen-vl/"><div class="blog-card-meta">Jan 24, 2024<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Introducing Qwen-VL</h2><p class="blog-card-desc">Along with the rapid development of our large language model Qwen, we leveraged Qwen’s capabilities and unified multimodal pretraining to address the limitations of multimodal models in generalization, and we opensourced multimodal model Qwen-VL in Sep. 2023. Recently, the Qwen-VL series has undergo</p></a><a class="blog-card" data-upstream="" href="/blog/qwen/"><div class="blog-card-meta">Jan 22, 2024<!-- --> <!-- -->·<!-- --> <!-- -->6<!-- --> min read</div><h2 class="blog-card-title">Introducing Qwen</h2><p class="blog-card-desc">4 months after our first release of Qwen-7B, which is the starting point of our opensource journey of large language models (LLM), we now provide an introduction to the Qwen series to give you a whole picture of our work as well as our objectives. Below are important links to our opensource projects</p></a><a class="blog-card" href="/blog/agent-nfts/"><div class="blog-card-meta">Dec 17, 2023<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Agent NFTs: Ownership and Identity for AI Agents</h2><p class="blog-card-desc">Introducing Agent NFTs, a framework for giving AI agents persistent identity and enabling ownership of their capabilities.</p><div class="blog-card-tags"><span class="blog-tag">Agents</span><span class="blog-tag">NFTs</span><span class="blog-tag">Blockchain</span></div></a><a class="blog-card" href="/blog/training-gym/"><div class="blog-card-meta">Sep 10, 2023<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Training Gym: A Platform for Open Model Development</h2><p class="blog-card-desc">Announcing Training Gym, our open platform for collaborative large model training.</p><div class="blog-card-tags"><span class="blog-tag">Infrastructure</span><span class="blog-tag">Training</span><span class="blog-tag">Open Source</span></div></a><a class="blog-card" href="/blog/proof-of-ai/"><div class="blog-card-meta">Jun 25, 2023<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Proof of AI: Verifiable Machine Learning on Chain</h2><p class="blog-card-desc">How we&#x27;re bringing cryptographic verification to AI inference, enabling trustless machine learning.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Blockchain</span><span class="blog-tag">Verification</span></div></a><a class="blog-card" href="/blog/zen-reranker/"><div class="blog-card-meta">Mar 12, 2023<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Zen Reranker: Two-Stage Retrieval Done Right</h2><p class="blog-card-desc">Introducing the Zen Reranker, a cross-encoder model that dramatically improves retrieval quality in two-stage pipelines.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Retrieval</span><span class="blog-tag">Reranking</span></div></a><a class="blog-card" data-upstream="" href="/blog/ofasys/"><div class="blog-card-meta">Dec 27, 2022<!-- --> <!-- -->·<!-- --> <!-- -->6<!-- --> min read</div><h2 class="blog-card-title">OFASys: Enabling Multitask Learning with One Line of Code! </h2><p class="blog-card-desc">Generalist Models are hot! We all see an opportunity towards a real generalist model by multimodal multitask learning. We previously release an opensourced unified multimodal pretrained model OFA for this goal. However, we actually met a lot of difficulties in our implementation. For example, it is</p></a><a class="blog-card" data-upstream="" href="/blog/chinese-clip/"><div class="blog-card-meta">Dec 23, 2022<!-- --> <!-- -->·<!-- --> <!-- -->5<!-- --> min read</div><h2 class="blog-card-title">Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese</h2><p class="blog-card-desc">CLIP[^1] is a phenomenal playmaker in vision and multimodal representation learning. It plays not only as a foundation model but also a bridge between vision and language. It has triggered a series of research in different fields, especially text-to-image generation. However, we find that there is a</p></a><a class="blog-card" href="/blog/7680-dim-embeddings/"><div class="blog-card-meta">Dec 4, 2022<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">7680-Dimensional Embeddings: More Dimensions, Better Retrieval</h2><p class="blog-card-desc">Why we trained embedding models with 7680 dimensions and what we learned about the relationship between dimensionality and retrieval quality.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Embeddings</span><span class="blog-tag">Retrieval</span></div></a><a class="blog-card" href="/blog/embedding-spaces-7680-dimensions/"><div class="blog-card-meta">Dec 4, 2022<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">Embedding Spaces at 7680 Dimensions</h2><p class="blog-card-desc">Exploring high-dimensional embedding spaces for semantic search and retrieval.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Embeddings</span><span class="blog-tag">Retrieval</span></div></a><a class="blog-card" data-upstream="" href="/blog/ofa/"><div class="blog-card-meta">Nov 13, 2022<!-- --> <!-- -->·<!-- --> <!-- -->10<!-- --> min read</div><h2 class="blog-card-title">OFA: Towards Building a One-For-All Model</h2><p class="blog-card-desc">2022 is a year of generalist models! With the bloom of multimodal pretraining, especially the unified model, we have witnessed the opportunity to building a generalist model that is capable of processing tasks of different modalities or multi-modalities! Thus, we propose OFA[^1], namely One-For-All,</p></a><a class="blog-card" href="/blog/grpo/"><div class="blog-card-meta">Sep 18, 2022<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">GRPO: Group Relative Policy Optimization</h2><p class="blog-card-desc">Introducing GRPO, a new approach to reinforcement learning from human feedback that improves sample efficiency and alignment stability.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Alignment</span><span class="blog-tag">Training</span></div></a><a class="blog-card" href="/blog/grpo-group-relative-policy-optimization/"><div class="blog-card-meta">Sep 17, 2022<!-- --> <!-- -->·<!-- --> <!-- -->4<!-- --> min read</div><h2 class="blog-card-title">GRPO: Group Relative Policy Optimization</h2><p class="blog-card-desc">A companion post to our GRPO paper, explaining group relative policy optimization for language model alignment.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">RLHF</span><span class="blog-tag">GRPO</span></div></a><a class="blog-card" href="/blog/federated-learning-without-compromise/"><div class="blog-card-meta">May 29, 2022<!-- --> <!-- -->·<!-- --> <!-- -->2<!-- --> min read</div><h2 class="blog-card-title">Federated Learning Without Compromise</h2><p class="blog-card-desc">Privacy-preserving machine learning that maintains model quality through novel aggregation protocols.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Federated Learning</span><span class="blog-tag">Privacy</span></div></a><a class="blog-card" href="/blog/federated-learning/"><div class="blog-card-meta">May 8, 2022<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Federated Learning for Open AI</h2><p class="blog-card-desc">How federated learning enables collaborative model training while preserving data privacy.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Training</span><span class="blog-tag">Privacy</span></div></a><a class="blog-card" href="/blog/experience-ledgers/"><div class="blog-card-meta">Feb 13, 2022<!-- --> <!-- -->·<!-- --> <!-- -->3<!-- --> min read</div><h2 class="blog-card-title">Experience Ledgers: Persistent Memory for AI Agents</h2><p class="blog-card-desc">Introducing experience ledgers, a framework for giving AI agents persistent, verifiable memory.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Agents</span></div></a><a class="blog-card" href="/blog/case-for-decentralized-science/"><div class="blog-card-meta">Nov 7, 2021<!-- --> <!-- -->·<!-- --> <!-- -->2<!-- --> min read</div><h2 class="blog-card-title">The Case for Decentralized Science</h2><p class="blog-card-desc">A manifesto for decentralized science (DeSci) and its application to AI research.</p><div class="blog-card-tags"><span class="blog-tag">DeSci</span><span class="blog-tag">Vision</span><span class="blog-tag">Governance</span></div></a><a class="blog-card" href="/blog/case-for-desci/"><div class="blog-card-meta">Nov 7, 2021<!-- --> <!-- -->·<!-- --> <!-- -->2<!-- --> min read</div><h2 class="blog-card-title">The Case for Decentralized Science</h2><p class="blog-card-desc">Why scientific research needs decentralization, and how blockchain can help.</p><div class="blog-card-tags"><span class="blog-tag">DeSci</span><span class="blog-tag">Vision</span></div></a><a class="blog-card" href="/blog/training-llms-collective-intelligence/"><div class="blog-card-meta">Jul 21, 2021<!-- --> <!-- -->·<!-- --> <!-- -->2<!-- --> min read</div><h2 class="blog-card-title">Training LLMs on Collective Intelligence</h2><p class="blog-card-desc">How we&#x27;re approaching training data curation to capture humanity&#x27;s collective intelligence.</p><div class="blog-card-tags"><span class="blog-tag">Research</span><span class="blog-tag">Training</span></div></a><a class="blog-card" href="/blog/introducing-zen/"><div class="blog-card-meta">Mar 14, 2021<!-- --> <!-- -->·<!-- --> <!-- -->1<!-- --> min read</div><h2 class="blog-card-title">Introducing Zen: Open AI for the Open Web</h2><p class="blog-card-desc">We&#x27;re launching Zen, an open research initiative to build AI that serves everyone.</p><div class="blog-card-tags"><span class="blog-tag">Announcement</span><span class="blog-tag">Vision</span></div></a></div></div></section></main><!--$--><!--/$--><footer><div class="container"><div class="footer-content"><div class="footer-section"><h4>Zen LM</h4><p>Open models from Zoo Labs Foundation, a 501(c)(3) non-profit, for agentic coding on your own machine and marketing work. Served on api.hanzo.ai.</p></div><div class="footer-section"><h4>Zen 6</h4><ul><li><a href="/models/#zen6">Zen 6</a></li><li><a href="/models/#zen6">Zen 6 Flash</a></li><li><a href="https://hanzo.ai/research-access" target="_blank" rel="noopener noreferrer">Zen 7 research preview</a></li></ul></div><div class="footer-section"><h4>Earlier Generations</h4><ul><li><a href="/models/#zen5">Zen 5</a></li><li><a href="/models/#zen4">Zen 4</a></li><li><a href="/models/#zen3">Zen 3</a></li></ul></div><div class="footer-section"><h4>Resources</h4><ul><li><a href="/datasets/">Training Data</a></li><li><a href="https://huggingface.co/zenlm" target="_blank" rel="noopener noreferrer">HuggingFace</a></li><li><a href="https://github.com/zenlm" target="_blank" rel="noopener noreferrer">GitHub</a></li><li><a href="/research/">Research Papers</a></li><li><a href="https://api.hanzo.ai" target="_blank" rel="noopener noreferrer">Zen API</a></li></ul></div></div><div class="footer-bottom"><p>© <!-- -->2026<!-- --> Zoo Labs Foundation, a 501(c)(3) non-profit. Zen LM open models.</p></div></div></footer><script src="/_next/static/chunks/webpack-f43ad49c79db3299.js" id="_R_" async=""></script><script>(self.__next_f=self.__next_f||[]).push([0])</script><script>self.__next_f.push([1,"1:\"$Sreact.fragment\"\n2:I[6529,[\"619\",\"static/chunks/619-ba102abea3e3d0e4.js\",\"177\",\"static/chunks/app/layout-425ca81883167c16.js\"],\"default\"]\n3:I[9766,[],\"\"]\n4:I[8924,[],\"\"]\n5:I[2619,[\"619\",\"static/chunks/619-ba102abea3e3d0e4.js\",\"831\",\"static/chunks/app/blog/page-f19d17b0f66e456e.js\"],\"\"]\n6:I[7405,[\"619\",\"static/chunks/619-ba102abea3e3d0e4.js\",\"177\",\"static/chunks/app/layout-425ca81883167c16.js\"],\"default\"]\na:I[7150,[],\"\"]\n:HL[\"/_next/static/css/1d1f6bc532e5f43f.css\",\"style\"]\n"])</script><script>self.__next_f.push([1,"0:{\"P\":null,\"b\":\"-BoRG_r6Z3v9udL0lI0J_\",\"p\":\"\",\"c\":[\"\",\"blog\",\"\"],\"i\":false,\"f\":[[[\"\",{\"children\":[\"blog\",{\"children\":[\"__PAGE__\",{}]}]},\"$undefined\",\"$undefined\",true],[\"\",[\"$\",\"$1\",\"c\",{\"children\":[[[\"$\",\"link\",\"0\",{\"rel\":\"stylesheet\",\"href\":\"/_next/static/css/1d1f6bc532e5f43f.css\",\"precedence\":\"next\",\"crossOrigin\":\"$undefined\",\"nonce\":\"$undefined\"}]],[\"$\",\"html\",null,{\"lang\":\"en\",\"children\":[[\"$\",\"head\",null,{\"children\":[[\"$\",\"link\",null,{\"rel\":\"icon\",\"type\":\"image/svg+xml\",\"href\":\"/favicon.svg\"}],[\"$\",\"link\",null,{\"rel\":\"alternate icon\",\"href\":\"/favicon.png\"}]]}],[\"$\",\"body\",null,{\"children\":[[\"$\",\"$L2\",null,{}],[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":[[[\"$\",\"title\",null,{\"children\":\"404: This page could not be found.\"}],[\"$\",\"div\",null,{\"style\":{\"fontFamily\":\"system-ui,\\\"Segoe UI\\\",Roboto,Helvetica,Arial,sans-serif,\\\"Apple Color Emoji\\\",\\\"Segoe UI Emoji\\\"\",\"height\":\"100vh\",\"textAlign\":\"center\",\"display\":\"flex\",\"flexDirection\":\"column\",\"alignItems\":\"center\",\"justifyContent\":\"center\"},\"children\":[\"$\",\"div\",null,{\"children\":[[\"$\",\"style\",null,{\"dangerouslySetInnerHTML\":{\"__html\":\"body{color:#000;background:#fff;margin:0}.next-error-h1{border-right:1px solid rgba(0,0,0,.3)}@media (prefers-color-scheme:dark){body{color:#fff;background:#000}.next-error-h1{border-right:1px solid rgba(255,255,255,.3)}}\"}}],[\"$\",\"h1\",null,{\"className\":\"next-error-h1\",\"style\":{\"display\":\"inline-block\",\"margin\":\"0 20px 0 0\",\"padding\":\"0 23px 0 0\",\"fontSize\":24,\"fontWeight\":500,\"verticalAlign\":\"top\",\"lineHeight\":\"49px\"},\"children\":404}],[\"$\",\"div\",null,{\"style\":{\"display\":\"inline-block\"},\"children\":[\"$\",\"h2\",null,{\"style\":{\"fontSize\":14,\"fontWeight\":400,\"lineHeight\":\"49px\",\"margin\":0},\"children\":\"This page could not be found.\"}]}]]}]}]],[]],\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}],[\"$\",\"footer\",null,{\"children\":[\"$\",\"div\",null,{\"className\":\"container\",\"children\":[[\"$\",\"div\",null,{\"className\":\"footer-content\",\"children\":[[\"$\",\"div\",null,{\"className\":\"footer-section\",\"children\":[[\"$\",\"h4\",null,{\"children\":\"Zen LM\"}],[\"$\",\"p\",null,{\"children\":\"Open models from Zoo Labs Foundation, a 501(c)(3) non-profit, for agentic coding on your own machine and marketing work. Served on api.hanzo.ai.\"}]]}],[\"$\",\"div\",null,{\"className\":\"footer-section\",\"children\":[[\"$\",\"h4\",null,{\"children\":\"Zen 6\"}],[\"$\",\"ul\",null,{\"children\":[[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/models#zen6\",\"children\":\"Zen 6\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/models#zen6\",\"children\":\"Zen 6 Flash\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"a\",null,{\"href\":\"https://hanzo.ai/research-access\",\"target\":\"_blank\",\"rel\":\"noopener noreferrer\",\"children\":\"Zen 7 research preview\"}]}]]}]]}],[\"$\",\"div\",null,{\"className\":\"footer-section\",\"children\":[[\"$\",\"h4\",null,{\"children\":\"Earlier Generations\"}],[\"$\",\"ul\",null,{\"children\":[[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/models#zen5\",\"children\":\"Zen 5\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/models#zen4\",\"children\":\"Zen 4\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/models#zen3\",\"children\":\"Zen 3\"}]}]]}]]}],[\"$\",\"div\",null,{\"className\":\"footer-section\",\"children\":[[\"$\",\"h4\",null,{\"children\":\"Resources\"}],[\"$\",\"ul\",null,{\"children\":[[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/datasets\",\"children\":\"Training Data\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"a\",null,{\"href\":\"https://huggingface.co/zenlm\",\"target\":\"_blank\",\"rel\":\"noopener noreferrer\",\"children\":\"HuggingFace\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"a\",null,{\"href\":\"https://github.com/zenlm\",\"target\":\"_blank\",\"rel\":\"noopener noreferrer\",\"children\":\"GitHub\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"$L5\",null,{\"href\":\"/research\",\"children\":\"Research Papers\"}]}],[\"$\",\"li\",null,{\"children\":[\"$\",\"a\",null,{\"href\":\"https://api.hanzo.ai\",\"target\":\"_blank\",\"rel\":\"noopener noreferrer\",\"children\":\"Zen API\"}]}]]}]]}]]}],[\"$\",\"div\",null,{\"className\":\"footer-bottom\",\"children\":[\"$\",\"p\",null,{\"children\":[\"© \",2026,\" Zoo Labs Foundation, a 501(c)(3) non-profit. Zen LM open models.\"]}]}]]}]}],[\"$\",\"$L6\",null,{}],[\"$\",\"script\",null,{\"src\":\"/assets/js/main.js\",\"async\":true}]]}]]}]]}],{\"children\":[\"blog\",[\"$\",\"$1\",\"c\",{\"children\":[null,[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":\"$undefined\",\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}]]}],{\"children\":[\"__PAGE__\",[\"$\",\"$1\",\"c\",{\"children\":[[\"$\",\"main\",null,{\"children\":[[\"$\",\"section\",null,{\"className\":\"hero\",\"children\":[\"$\",\"div\",null,{\"className\":\"container\",\"children\":[[\"$\",\"h1\",null,{\"className\":\"hero-title\",\"children\":\"Blog\"}],[\"$\",\"p\",null,{\"className\":\"hero-subtitle\",\"children\":\"Research notes, deep dives, and releases\"}],[\"$\",\"p\",null,{\"className\":\"hero-description\",\"children\":[71,\" posts on Zen architecture, training, and the open-model ecosystem.\"]}]]}]}],\"$L7\"]}],null,\"$L8\"]}],{},null,false]},null,false]},null,false],\"$L9\",false]],\"m\":\"$undefined\",\"G\":[\"$a\",[]],\"s\":false,\"S\":true}\n"])</script><script>self.__next_f.push([1,"4e:I[4431,[],\"OutletBoundary\"]\n50:I[5278,[],\"AsyncMetadataOutlet\"]\n52:I[4431,[],\"ViewportBoundary\"]\n54:I[4431,[],\"MetadataBoundary\"]\n55:\"$Sreact.suspense\"\n"])</script><script>self.__next_f.push([1,"7:[\"$\",\"section\",null,{\"className\":\"catalog-section\",\"children\":[\"$\",\"div\",null,{\"className\":\"container\",\"children\":[\"$\",\"div\",null,{\"className\":\"blog-list\",\"children\":[[\"$\",\"$L5\",\"bitdelta-behavioral-compression\",{\"className\":\"blog-card\",\"href\":\"/blog/bitdelta-behavioral-compression\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 27, 2026\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How BitDelta (arXiv:2402.10193) compresses fine-tuned behavioral deltas to 1-bit precision, enabling the full Zen model family — nano through ultra — to share a single GPU cluster.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Quantization\",{\"className\":\"blog-tag\",\"children\":\"Quantization\"}],[\"$\",\"span\",\"BitDelta\",{\"className\":\"blog-tag\",\"children\":\"BitDelta\"}]]}]]}],[\"$\",\"$L5\",\"continual-learning-sure-opcm\",{\"className\":\"blog-card\",\"href\":\"/blog/continual-learning-sure-opcm\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 27, 2026\",\" \",\"·\",\" \",8,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"SuRe + OPCM: Production-Grade Continual Learning for Open Models\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Deep dive on Surprise-Driven Prioritized Replay (SuRe) and Orthogonal Projection Continual Merging (OPCM) — the two SOTA techniques we use for catastrophic-forgetting-free LLM adaptation in the Zen model family.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Continual Learning\",{\"className\":\"blog-tag\",\"children\":\"Continual Learning\"}],[\"$\",\"span\",\"SuRe\",{\"className\":\"blog-tag\",\"children\":\"SuRe\"}]]}]]}],[\"$\",\"$L5\",\"drop-upcycling-zen-mode\",{\"className\":\"blog-card\",\"href\":\"/blog/drop-upcycling-zen-mode\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 27, 2026\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Drop-Upcycling and the Birth of Zen MoDE Architecture\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How Drop-Upcycling (arXiv:2502.19261) transforms dense checkpoints into MoE models at 1/4 training cost, and how it shapes Zen MoDE — our Mixture of Distilled Experts architecture.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"MoE\",{\"className\":\"blog-tag\",\"children\":\"MoE\"}],[\"$\",\"span\",\"Architecture\",{\"className\":\"blog-tag\",\"children\":\"Architecture\"}]]}]]}],[\"$\",\"$L5\",\"gt-qlora-moe-abliteration\",{\"className\":\"blog-card\",\"href\":\"/blog/gt-qlora-moe-abliteration\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 27, 2026\",\" \",\"·\",\" \",8,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"GT-QLoRA: Uncensoring Trillion-Parameter MoE Models\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Why standard abliteration techniques fail on Mixture-of-Experts models, and how Gate-Targeted QLoRA solves the expert routing problem at 1 trillion parameters.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Abliteration\",{\"className\":\"blog-tag\",\"children\":\"Abliteration\"}],[\"$\",\"span\",\"GT-QLoRA\",{\"className\":\"blog-tag\",\"children\":\"GT-QLoRA\"}]]}]]}],[\"$\",\"$L5\",\"zen4-ultra\",{\"className\":\"blog-card\",\"href\":\"/blog/zen4-ultra\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 19, 2026\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Zen4 Ultra: 480B Parameters, 1M Token Context\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Zen4 Ultra is our most capable model: 480B total parameters, 35B active per token, 1M token context window. Benchmark results and use cases.\"}],\"$Lb\"]}],\"$Lc\",\"$Ld\",\"$Le\",\"$Lf\",\"$L10\",\"$L11\",\"$L12\",\"$L13\",\"$L14\",\"$L15\",\"$L16\",\"$L17\",\"$L18\",\"$L19\",\"$L1a\",\"$L1b\",\"$L1c\",\"$L1d\",\"$L1e\",\"$L1f\",\"$L20\",\"$L21\",\"$L22\",\"$L23\",\"$L24\",\"$L25\",\"$L26\",\"$L27\",\"$L28\",\"$L29\",\"$L2a\",\"$L2b\",\"$L2c\",\"$L2d\",\"$L2e\",\"$L2f\",\"$L30\",\"$L31\",\"$L32\",\"$L33\",\"$L34\",\"$L35\",\"$L36\",\"$L37\",\"$L38\",\"$L39\",\"$L3a\",\"$L3b\",\"$L3c\",\"$L3d\",\"$L3e\",\"$L3f\",\"$L40\",\"$L41\",\"$L42\",\"$L43\",\"$L44\",\"$L45\",\"$L46\",\"$L47\",\"$L48\",\"$L49\",\"$L4a\",\"$L4b\",\"$L4c\",\"$L4d\"]}]}]}]\n"])</script><script>self.__next_f.push([1,"8:[\"$\",\"$L4e\",null,{\"children\":[\"$L4f\",[\"$\",\"$L50\",null,{\"promise\":\"$@51\"}]]}]\n9:[\"$\",\"$1\",\"h\",{\"children\":[null,[[\"$\",\"$L52\",null,{\"children\":\"$L53\"}],null],[\"$\",\"$L54\",null,{\"children\":[\"$\",\"div\",null,{\"hidden\":true,\"children\":[\"$\",\"$55\",null,{\"fallback\":null,\"children\":\"$L56\"}]}]}]]}]\n"])</script><script>self.__next_f.push([1,"b:[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Models\",{\"className\":\"blog-tag\",\"children\":\"Models\"}],[\"$\",\"span\",\"Zen4\",{\"className\":\"blog-tag\",\"children\":\"Zen4\"}],[\"$\",\"span\",\"Flagship\",{\"className\":\"blog-tag\",\"children\":\"Flagship\"}]]}]\n"])</script><script>self.__next_f.push([1,"c:[\"$\",\"$L5\",\"zen-mode-architecture\",{\"className\":\"blog-card\",\"href\":\"/blog/zen-mode-architecture\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 15, 2026\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Zen MoDE: Mixture of Distilled Experts\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"A technical deep dive into Zen MoDE — Mixture of Distilled Experts — the architecture underlying all Zen models.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Architecture\",{\"className\":\"blog-tag\",\"children\":\"Architecture\"}],[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"MoE\",{\"className\":\"blog-tag\",\"children\":\"MoE\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"d:[\"$\",\"$L5\",\"zen-lm-launch\",{\"className\":\"blog-card\",\"href\":\"/blog/zen-lm-launch\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 14, 2026\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Zen LM: Open Frontier Models from Hanzo AI and Zoo Labs\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Announcing the Zen model family: 94+ open models built on Zen MoDE architecture, co-developed by Hanzo AI and Zoo Labs Foundation.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Announcement\",{\"className\":\"blog-tag\",\"children\":\"Announcement\"}],[\"$\",\"span\",\"Models\",{\"className\":\"blog-tag\",\"children\":\"Models\"}],[\"$\",\"span\",\"Zen\",{\"className\":\"blog-tag\",\"children\":\"Zen\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"e:[\"$\",\"$L5\",\"qwen3guard\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen3guard\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 22, 2025\",\" \",\"·\",\" \",8,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen3Guard: Real-time Safety for Your Token Stream\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We are excited to introduce Qwen3Guard, the first safety guardrail model in the Qwen family. Built upon the powerful Qwen3 foundation models and fine-tuned specifically for safety classificatoin, Qwen3Guard ensures responsible AI interactions by delivering precise safety detection for both prompts a\"}],false]}]\n"])</script><script>self.__next_f.push([1,"f:[\"$\",\"$L5\",\"qwen-image-edit\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-image-edit\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 18, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen-Image-Edit: Image Editing with Higher Quality and Efficiency\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We are excited to introduce Qwen-Image-Edit, the image editing version of Qwen-Image. Built upon our 20B Qwen-Image model, Qwen-Image-Edit successfully extends Qwen-Image's unique text rendering capabilities to image editing tasks, enabling precise text editing. Furthermore, Qwen-Image-Edit simultan\"}],false]}]\n"])</script><script>self.__next_f.push([1,"10:[\"$\",\"$L5\",\"qwen-image\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-image\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 3, 2025\",\" \",\"·\",\" \",12,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen-Image: Crafting with Native Text Rendering\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We are thrilled to release **Qwen-Image**, a 20B MMDiT image foundation model that achieves significant advances in complex text rendering and precise image editing. To try the latest model, feel free to visit [Qwen Chat](https://chat.qwenlm.ai) and choose “Image Generation”.\"}],false]}]\n"])</script><script>self.__next_f.push([1,"11:[\"$\",\"$L5\",\"gspo\",{\"className\":\"blog-card\",\"href\":\"/blog/gspo\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jul 26, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"GSPO: Towards Scalable Reinforcement Learning for Language Models\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Reinforcement Learning (RL) has emerged as a pivotal paradigm for scaling language models and enhancing their deep reasoning and problem-solving capabilities. To scale RL, the foremost prerequisite is maintaining stable and robust training dynamics. However, we observe that existing RL algorithms (s\"}],false]}]\n"])</script><script>self.__next_f.push([1,"12:[\"$\",\"$L5\",\"qwen-mt\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-mt\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jul 23, 2025\",\" \",\"·\",\" \",9,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen-MT: Where Speed Meets Smart Translation\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Here we introduce the latest update of Qwen-MT (qwen-mt-turbo) via [Qwen API](https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?type=model\u0026url=https://www.alibabacloud.com/help/en/doc-detail/2840914_2.html\u0026renderType=component\u0026modelId=qwen-mt-turbo). This update builds upon the powerful Qw\"}],false]}]\n"])</script><script>self.__next_f.push([1,"13:[\"$\",\"$L5\",\"qwen3-coder\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen3-coder\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jul 21, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen3-Coder: Agentic Coding in the World\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Today, we're announcing Qwen3-Coder, our most agentic code model to date. Qwen3-Coder is available in multiple sizes, but we're excited to introduce its most powerful variant first: Qwen3-Coder-480B-A35B-Instruct — a 480B-parameter Mixture-of-Experts model with 35B active parameters which supports t\"}],false]}]\n"])</script><script>self.__next_f.push([1,"14:[\"$\",\"$L5\",\"qwen-tts\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-tts\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 26, 2025\",\" \",\"·\",\" \",6,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Time to Speak Some Dialects, Qwen-TTS!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Here we introduce the latest update of **Qwen-TTS** (`qwen-tts-latest` or `qwen-tts-2025-05-22`) through [Qwen API](https://help.aliyun.com/zh/model-studio/qwen-tts) . Trained on a large-scale dataset encompassing over millions of hours of speech, Qwen-TTS achieves human-level naturalness and expres\"}],false]}]\n"])</script><script>self.__next_f.push([1,"15:[\"$\",\"$L5\",\"qwen-vlo\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-vlo\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 25, 2025\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen VLo: From \\\"Understanding\\\" the World to \\\"Depicting\\\" It\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"The evolution of multimodal large models is continually pushing the boundaries of what we believe technology can achieve. From the initial QwenVL to the latest Qwen2.5 VL, we have made progress in enhancing the model's ability to understand image content. Today, we are excited to introduce a new model,\"}],false]}]\n"])</script><script>self.__next_f.push([1,"16:[\"$\",\"$L5\",\"qwen3-embedding\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen3-embedding\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 4, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"{/* {{\u003c video src=\\\"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/qwen2-main-video.m4v\\\" width=\\\"100%\\\" alt=\\\"Qwen2.5 Main Video\\\" autoplay=true loop=true controls=false muted=true playsinline=true \u003e}} */}\"}],false]}]\n17:[\"$\",\"$L5\",\"qwen3\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen3\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Apr 28, 2025\",\" \",\"·\",\" \",11,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen3: Think Deeper, Act Faster\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Today, we are excited to announce the release of **Qwen3**, the latest addition to the Qwen family of large language models. Our flagship model, **Qwen3-235B-A22B**, achieves competitive results in benchmark evaluations of coding, math, general capabilities, etc., when compared to other top-tier mod\"}],false]}]\n18:[\"$\",\"$L5\",\"qvq-max-preview\",{\"className\":\"blog-card\",\"href\":\"/blog/qvq-max-preview\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 27, 2025\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"QVQ-Max: Think with Evidence\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Last December, we launched QVQ-72B-Preview as an exploratory model, but it had many issues. Today, we are officially releasing the first version of QVQ-Max, our visual reasoning model. This model can not only \\\"understand\\\" the content in images and videos but also analyze and reason with this informa\"}],false]}]\n"])</script><script>self.__next_f.push([1,"19:[\"$\",\"$L5\",\"qwen2.5-omni\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-omni\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 26, 2025\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5 Omni: See, Hear, Talk, Write, Do It All!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We release **Qwen2.5-Omni**, the new flagship end-to-end multimodal model in the Qwen series. Designed for comprehensive multimodal perception, it seamlessly processes diverse inputs including text, images, audio, and video, while delivering real-time streaming responses through both text generation and\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1a:[\"$\",\"$L5\",\"qwen2.5-vl-32b\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-vl-32b\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 23, 2025\",\" \",\"·\",\" \",2,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-VL-32B: Smarter and Lighter\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"At the end of January this year, we launched the Qwen2.5-VL series of models, which received widespread attention and positive feedback from the community. Building on the Qwen2.5-VL series, we continued to optimize the model using reinforcement learning and open-sourced the new VL model with the beloved 32\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1b:[\"$\",\"$L5\",\"qwq-32b\",{\"className\":\"blog-card\",\"href\":\"/blog/qwq-32b\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 5, 2025\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"QwQ-32B: Embracing the Power of Reinforcement Learning\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Scaling Reinforcement Learning (RL) has the potential to enhance model performance beyond conventional pretraining and post-training methods. Recent studies have demonstrated that RL can significantly improve the reasoning capabilities of models. For instance, DeepSeek R1 has achieved state-of-the-a\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1c:[\"$\",\"$L5\",\"qwq-max-preview\",{\"className\":\"blog-card\",\"href\":\"/blog/qwq-max-preview\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 24, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"\u003cthink\u003e...\u003c/think\u003e QwQ-Max-Preview\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"This is a blog created by QwQ-Max-Preview. We hope you enjoy it!\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1d:[\"$\",\"$L5\",\"qwen2.5-max\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-max\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 27, 2025\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"It is widely recognized that continuously scaling both data size and model size can lead to significant improvements in model intelligence. However, the research and industry community has limited experience in effectively scaling extremely large models, whether they are dense or Mixture-of-Expert (\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1e:[\"$\",\"$L5\",\"qwen2.5-1m\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-1m\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 26, 2025\",\" \",\"·\",\" \",9,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Two months after upgrading [Qwen2.5-Turbo](../qwen2.5-turbo) to support context length up to one million tokens, we are back with the open-source Qwen2.5-1M models and the corresponding inference framework support. Here's what you can expect from this release:\"}],false]}]\n"])</script><script>self.__next_f.push([1,"1f:[\"$\",\"$L5\",\"qwen2.5-vl\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-vl\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 25, 2025\",\" \",\"·\",\" \",6,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We release **Qwen2.5-VL**, the new flagship vision-language model of Qwen and also a significant leap from the previous Qwen2-VL. To try the latest model, feel free to visit [Qwen Chat](https://chat.qwenlm.ai) and choose Qwen2.5-VL-72B-Instruct. Also, we open both base and instruct models in 3 sizes, includin\"}],false]}]\n"])</script><script>self.__next_f.push([1,"20:[\"$\",\"$L5\",\"global-load-balance\",{\"className\":\"blog-card\",\"href\":\"/blog/global-load-balance\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 20, 2025\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Global-batch load balance almost free lunch to improve your MoE LLM training\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"The Mixture-of-Experts (MoEs) architecture has become a popular model-parameter-scale-up technique. Typically, one MoE layer consists of a router (often parameterized as one single Linear layer) and a group of experts (for transformer-based models, each expert is one feedforward layer). Given an inp\"}],false]}]\n"])</script><script>self.__next_f.push([1,"21:[\"$\",\"$L5\",\"qwen2.5-math-prm\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-math-prm\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 13, 2025\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Towards Effective Process Supervision in Mathematical Reasoning\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"In recent years, Large Language Models (LLMs) have made remarkable advances in mathematical reasoning, yet they can make mistakes, such as miscalculations or logical errors, leading to wrong conclusions. Moreover, even when achieving correct final answers, these powerful models can still regularly m\"}],false]}]\n"])</script><script>self.__next_f.push([1,"22:[\"$\",\"$L5\",\"zen-3\",{\"className\":\"blog-card\",\"href\":\"/blog/zen-3\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 12, 2025\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Zen 3.0: Our Third-Generation Open Model Family\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Announcing Zen 3.0, our most capable open model family yet.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Announcement\",{\"className\":\"blog-tag\",\"children\":\"Announcement\"}],[\"$\",\"span\",\"Models\",{\"className\":\"blog-tag\",\"children\":\"Models\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"23:[\"$\",\"$L5\",\"qvq-72b-preview\",{\"className\":\"blog-card\",\"href\":\"/blog/qvq-72b-preview\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 24, 2024\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"QVQ: To See the World with Wisdom\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Language and vision intertwine in the human mind, shaping how we perceive and understand the world around us. Our ability to reason is deeply rooted in both linguistic thought and visual memory - but what happens when we extend these capabilities to AI? Today's large language models have demonstrate\"}],false]}]\n"])</script><script>self.__next_f.push([1,"24:[\"$\",\"$L5\",\"qwq-32b-preview\",{\"className\":\"blog-card\",\"href\":\"/blog/qwq-32b-preview\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 27, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"QwQ: Reflect Deeply on the Boundaries of the Unknown\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"\u003ci style=\\\"color: grey;\\\"\u003eNote: This is the pronunciation of QwQ: /kwju:/ , similar to the word \\\"quill\\\".\u003c/i\u003e\"}],false]}]\n"])</script><script>self.__next_f.push([1,"25:[\"$\",\"$L5\",\"qwen2.5-turbo\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-turbo\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 14, 2024\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Extending the Context Length to 1M Tokens!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"After the release of Qwen2.5, we heard the community's demand for processing longer contexts. In recent months, we have made many optimizations for the model capabilities and inference performance of extremely long context. Today, we are proud to introduce the new Qwen2.5-Turbo version, which features:\"}],false]}]\n"])</script><script>self.__next_f.push([1,"26:[\"$\",\"$L5\",\"qwen2.5-coder-family\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-coder-family\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 11, 2024\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-Coder Series: Powerful, Diverse, Practical.\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"{/* {{\u003c figure src=\\\"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/Qwen2.5-Coder/coder-main.png#center\\\" width=\\\"70%\\\"\u003e}} */}\"}],false]}]\n27:[\"$\",\"$L5\",\"future-of-open-ai\",{\"className\":\"blog-card\",\"href\":\"/blog/future-of-open-ai\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 10, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"The Future of Open AI\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Reflections on where open AI development is heading and what it will take to get there.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Vision\",{\"className\":\"blog-tag\",\"children\":\"Vision\"}],[\"$\",\"span\",\"Open Source\",{\"className\":\"blog-tag\",\"children\":\"Open Source\"}]]}]]}]\n28:[\"$\",\"$L5\",\"qwen2.5\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 18, 2024\",\" \",\"·\",\" \",11,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5: A Party of Foundation Models!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"{/* {{\u003c video src=\\\"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/qwen2-main-video.m4v\\\" width=\\\"100%\\\" alt=\\\"Qwen2.5 Main Video\\\" autoplay=true loop=true controls=false muted=true playsinline=true \u003e}} */}\"}],false]}]\n"])</script><script>self.__next_f.push([1,"29:[\"$\",\"$L5\",\"qwen2.5-coder\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-coder\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 18, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-Coder: Code More, Learn More!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"In early April, we introduced CodeQwen1.5, which garnered significant attention from the community. Since then, we have been working to enhance the coding model. Today, we are excited to announce the release of the next generation of open-source coding models, **Qwen2.5-Coder**, and officially rename Co\"}],false]}]\n"])</script><script>self.__next_f.push([1,"2a:[\"$\",\"$L5\",\"qwen2.5-llm\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-llm\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 18, 2024\",\" \",\"·\",\" \",23,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-LLM: Extending the boundary of LLMs\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"In this blog, we delve into the details of our latest Qwen2.5 series language models. We have developed a range of decoder-only dense models, with seven of them open-sourced, spanning from 0.5B to 72B parameters. Our research indicates a significant interest among users in models within the 10-30B range\"}],false]}]\n"])</script><script>self.__next_f.push([1,"2b:[\"$\",\"$L5\",\"qwen2.5-math\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2.5-math\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 18, 2024\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2.5-Math: The world's leading open-sourced mathematical LLMs\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"\u003e \u003cdiv align=\\\"center\\\"\u003e \u003e \u003cb\u003e \u003e 🚨 Qwen2.5-Math mainly supports solving English and Chinese math problems through CoT and TIR. We do not recommend using this series of models for other tasks. \u003e \u003c/b\u003e \u003e \u003c/div\u003e\"}],false]}]\n2c:[\"$\",\"$L5\",\"qwen2-vl\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2-vl\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 28, 2024\",\" \",\"·\",\" \",8,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2-VL: To See the World More Clearly\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"After a year's relentless efforts, today we are thrilled to release **Qwen2-VL**! Qwen2-VL is the latest version of the vision language models based on **Qwen2** in the Qwen model familities. Compared with Qwen-VL, Qwen2-VL has the capabilities of:\"}],false]}]\n2d:[\"$\",\"$L5\",\"qwen2-audio\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2-audio\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 8, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen2-Audio: Chat with Your Voice!\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"To achieve the objective of building an AGI system, the model should be capable of understanding information from different modalities. Thanks to the rapid development of large language models, LLMs are now capable of understanding language and reasoning. Previously we have taken a step forward to e\"}],false]}]\n2e:[\"$\",\"$L5\",\"qwen2-math\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2-math\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",nul"])</script><script>self.__next_f.push([1,"l,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 7, 2024\",\" \",\"·\",\" \",30,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Qwen2-Math\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"\u003e \u003cdiv align=\\\"center\\\"\u003e \u003e \u003cb\u003e \u003e 🚨 This model mainly supports English. We will release bilingual (English and Chinese) math models soon. \u003e \u003c/b\u003e \u003e \u003c/div\u003e\"}],false]}]\n"])</script><script>self.__next_f.push([1,"2f:[\"$\",\"$L5\",\"decentralized-compute\",{\"className\":\"blog-card\",\"href\":\"/blog/decentralized-compute\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Aug 4, 2024\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Decentralized Compute for AI Training\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How we're building a decentralized compute network for training large AI models.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Infrastructure\",{\"className\":\"blog-tag\",\"children\":\"Infrastructure\"}],[\"$\",\"span\",\"Training\",{\"className\":\"blog-tag\",\"children\":\"Training\"}],[\"$\",\"span\",\"Decentralization\",{\"className\":\"blog-tag\",\"children\":\"Decentralization\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"30:[\"$\",\"$L5\",\"qwen2\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen2\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 6, 2024\",\" \",\"·\",\" \",25,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Hello Qwen2\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"After months of efforts, we are pleased to announce the evolution from Qwen1.5 to Qwen2. This time, we bring to you:\"}],false]}]\n31:[\"$\",\"$L5\",\"qwen-agent-2405\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-agent-2405\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 5, 2024\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Generalizing an LLM from 8k to 1M Context using Qwen-Agent\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"**TLDR:** We've created an agent using Qwen2 models with an 8k context size to understand documents with 1M tokens, surpassing RAG and native long-context models. This agent was also used to generate data for training new long-context Qwen models.\"}],false]}]\n"])</script><script>self.__next_f.push([1,"32:[\"$\",\"$L5\",\"zips-governance\",{\"className\":\"blog-card\",\"href\":\"/blog/zips-governance\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"May 19, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"ZIPs: Decentralized Governance for Open AI\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How Zoo Improvement Proposals enable community-driven governance of open AI development.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Governance\",{\"className\":\"blog-tag\",\"children\":\"Governance\"}],[\"$\",\"span\",\"ZIPs\",{\"className\":\"blog-tag\",\"children\":\"ZIPs\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"33:[\"$\",\"$L5\",\"qwen-max-0428\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-max-0428\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"May 10, 2024\",\" \",\"·\",\" \",1,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Notes on Qwen-Max-0428\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Previously, we opensourced a series of Qwen1.5 model ranging from 0.5 to 110 billion parameters. Now, we release a larger model, Qwen-Max-0428. Qwen-Max-0428 is an instruction-tuned model for chat service. Very recently, it is available via [Chatbot Arena](https://chat.lmsys.org/) and it has now bec\"}],false]}]\n"])</script><script>self.__next_f.push([1,"34:[\"$\",\"$L5\",\"qwen1.5-110b\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen1.5-110b\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Apr 24, 2024\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen1.5-110B: The First 100B+ Model of the Qwen1.5 Series\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Recently we have witnessed a burst of large-scale models with over 100 billion parameters in the opensource community. These models have demonstrated remarkable performance in both benchmark evaluation and chatbot arena. Today, we release the first 100B+ model of the Qwen1.5 series, Qwen1.5-110B, wh\"}],false]}]\n"])</script><script>self.__next_f.push([1,"35:[\"$\",\"$L5\",\"codeqwen1.5\",{\"className\":\"blog-card\",\"href\":\"/blog/codeqwen1.5\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Apr 15, 2024\",\" \",\"·\",\" \",7,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Code with CodeQwen1.5\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"The advent of advanced programming tools, which harnesses the power of large language models (LLMs), has significantly enhanced programmer productivity and accuracy. Notwithstanding these advancements, dominant coding assistants like Github Copilot, built upon proprietary LLMs, pose notable challeng\"}],false]}]\n"])</script><script>self.__next_f.push([1,"36:[\"$\",\"$L5\",\"qwen1.5-32b\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen1.5-32b\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Apr 1, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen1.5-32B: Fitting the Capstone of the Qwen1.5 Language Model Series\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"The open-source community has long sought a model that strikes an ideal balance between performance, efficiency, and memory footprint. Despite the emergence of cutting-edge models like Qwen1.5-72B and DBRX, the models have faced persistent challenges such as large memory consumption, slow inference\"}],false]}]\n"])</script><script>self.__next_f.push([1,"37:[\"$\",\"$L5\",\"qwen-moe\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-moe\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 27, 2024\",\" \",\"·\",\" \",8,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Since the surge in interest sparked by Mixtral, research on mixture-of-expert (MoE) models has gained significant momentum. Both researchers and practitioners are keenly interested in understanding how to effectively train such models and assessing their efficiency and effectiveness. Today, we intro\"}],false]}]\n"])</script><script>self.__next_f.push([1,"38:[\"$\",\"$L5\",\"zoo-foundation-launch\",{\"className\":\"blog-card\",\"href\":\"/blog/zoo-foundation-launch\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 11, 2024\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Announcing the Zoo Labs Foundation\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Today we formally launch Zoo Labs Foundation, an open research network dedicated to decentralized AI and decentralized science.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Announcement\",{\"className\":\"blog-tag\",\"children\":\"Announcement\"}],[\"$\",\"span\",\"Foundation\",{\"className\":\"blog-tag\",\"children\":\"Foundation\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"39:[\"$\",\"$L5\",\"qwen1.5\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen1.5\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 3, 2024\",\" \",\"·\",\" \",18,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Qwen1.5\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"In recent months, our focus has been on developing a \\\"good\\\" model while optimizing the developer experience. As we progress towards **Qwen1.5**, the next iteration in our Qwen series, this update arrives just before the Chinese New Year.\"}],false]}]\n3a:[\"$\",\"$L5\",\"qwen-vl\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen-vl\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 24, 2024\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Qwen-VL\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Along with the rapid development of our large language model Qwen, we leveraged Qwen’s capabilities and unified multimodal pretraining to address the limitations of multimodal models in generalization, and we opensourced multimodal model Qwen-VL in Sep. 2023. Recently, the Qwen-VL series has undergo\"}],false]}]\n3b:[\"$\",\"$L5\",\"qwen\",{\"className\":\"blog-card\",\"href\":\"/blog/qwen\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jan 22, 2024\",\" \",\"·\",\" \",6,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Qwen\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"4 months after our first release of Qwen-7B, which is the starting point of our opensource journey of large language models (LLM), we now provide an introduction to the Qwen series to give you a whole picture of our work as well as our objectives. Below are important links to our opensource projects\"}],false]}]\n"])</script><script>self.__next_f.push([1,"3c:[\"$\",\"$L5\",\"agent-nfts\",{\"className\":\"blog-card\",\"href\":\"/blog/agent-nfts\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 17, 2023\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Agent NFTs: Ownership and Identity for AI Agents\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Introducing Agent NFTs, a framework for giving AI agents persistent identity and enabling ownership of their capabilities.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Agents\",{\"className\":\"blog-tag\",\"children\":\"Agents\"}],[\"$\",\"span\",\"NFTs\",{\"className\":\"blog-tag\",\"children\":\"NFTs\"}],[\"$\",\"span\",\"Blockchain\",{\"className\":\"blog-tag\",\"children\":\"Blockchain\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"3d:[\"$\",\"$L5\",\"training-gym\",{\"className\":\"blog-card\",\"href\":\"/blog/training-gym\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 10, 2023\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Training Gym: A Platform for Open Model Development\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Announcing Training Gym, our open platform for collaborative large model training.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Infrastructure\",{\"className\":\"blog-tag\",\"children\":\"Infrastructure\"}],[\"$\",\"span\",\"Training\",{\"className\":\"blog-tag\",\"children\":\"Training\"}],[\"$\",\"span\",\"Open Source\",{\"className\":\"blog-tag\",\"children\":\"Open Source\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"3e:[\"$\",\"$L5\",\"proof-of-ai\",{\"className\":\"blog-card\",\"href\":\"/blog/proof-of-ai\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jun 25, 2023\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Proof of AI: Verifiable Machine Learning on Chain\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How we're bringing cryptographic verification to AI inference, enabling trustless machine learning.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Blockchain\",{\"className\":\"blog-tag\",\"children\":\"Blockchain\"}],[\"$\",\"span\",\"Verification\",{\"className\":\"blog-tag\",\"children\":\"Verification\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"3f:[\"$\",\"$L5\",\"zen-reranker\",{\"className\":\"blog-card\",\"href\":\"/blog/zen-reranker\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 12, 2023\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Zen Reranker: Two-Stage Retrieval Done Right\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Introducing the Zen Reranker, a cross-encoder model that dramatically improves retrieval quality in two-stage pipelines.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Retrieval\",{\"className\":\"blog-tag\",\"children\":\"Retrieval\"}],[\"$\",\"span\",\"Reranking\",{\"className\":\"blog-tag\",\"children\":\"Reranking\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"40:[\"$\",\"$L5\",\"ofasys\",{\"className\":\"blog-card\",\"href\":\"/blog/ofasys\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 27, 2022\",\" \",\"·\",\" \",6,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"OFASys: Enabling Multitask Learning with One Line of Code! \"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Generalist Models are hot! We all see an opportunity towards a real generalist model by multimodal multitask learning. We previously release an opensourced unified multimodal pretrained model OFA for this goal. However, we actually met a lot of difficulties in our implementation. For example, it is\"}],false]}]\n"])</script><script>self.__next_f.push([1,"41:[\"$\",\"$L5\",\"chinese-clip\",{\"className\":\"blog-card\",\"href\":\"/blog/chinese-clip\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 23, 2022\",\" \",\"·\",\" \",5,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"CLIP[^1] is a phenomenal playmaker in vision and multimodal representation learning. It plays not only as a foundation model but also a bridge between vision and language. It has triggered a series of research in different fields, especially text-to-image generation. However, we find that there is a\"}],false]}]\n"])</script><script>self.__next_f.push([1,"42:[\"$\",\"$L5\",\"7680-dim-embeddings\",{\"className\":\"blog-card\",\"href\":\"/blog/7680-dim-embeddings\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 4, 2022\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"7680-Dimensional Embeddings: More Dimensions, Better Retrieval\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Why we trained embedding models with 7680 dimensions and what we learned about the relationship between dimensionality and retrieval quality.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Embeddings\",{\"className\":\"blog-tag\",\"children\":\"Embeddings\"}],[\"$\",\"span\",\"Retrieval\",{\"className\":\"blog-tag\",\"children\":\"Retrieval\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"43:[\"$\",\"$L5\",\"embedding-spaces-7680-dimensions\",{\"className\":\"blog-card\",\"href\":\"/blog/embedding-spaces-7680-dimensions\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Dec 4, 2022\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Embedding Spaces at 7680 Dimensions\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Exploring high-dimensional embedding spaces for semantic search and retrieval.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Embeddings\",{\"className\":\"blog-tag\",\"children\":\"Embeddings\"}],[\"$\",\"span\",\"Retrieval\",{\"className\":\"blog-tag\",\"children\":\"Retrieval\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"44:[\"$\",\"$L5\",\"ofa\",{\"className\":\"blog-card\",\"href\":\"/blog/ofa\",\"data-upstream\":\"\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 13, 2022\",\" \",\"·\",\" \",10,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"OFA: Towards Building a One-For-All Model\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"2022 is a year of generalist models! With the bloom of multimodal pretraining, especially the unified model, we have witnessed the opportunity to building a generalist model that is capable of processing tasks of different modalities or multi-modalities! Thus, we propose OFA[^1], namely One-For-All,\"}],false]}]\n"])</script><script>self.__next_f.push([1,"45:[\"$\",\"$L5\",\"grpo\",{\"className\":\"blog-card\",\"href\":\"/blog/grpo\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 18, 2022\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"GRPO: Group Relative Policy Optimization\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Introducing GRPO, a new approach to reinforcement learning from human feedback that improves sample efficiency and alignment stability.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Alignment\",{\"className\":\"blog-tag\",\"children\":\"Alignment\"}],[\"$\",\"span\",\"Training\",{\"className\":\"blog-tag\",\"children\":\"Training\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"46:[\"$\",\"$L5\",\"grpo-group-relative-policy-optimization\",{\"className\":\"blog-card\",\"href\":\"/blog/grpo-group-relative-policy-optimization\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Sep 17, 2022\",\" \",\"·\",\" \",4,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"GRPO: Group Relative Policy Optimization\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"A companion post to our GRPO paper, explaining group relative policy optimization for language model alignment.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"RLHF\",{\"className\":\"blog-tag\",\"children\":\"RLHF\"}],[\"$\",\"span\",\"GRPO\",{\"className\":\"blog-tag\",\"children\":\"GRPO\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"47:[\"$\",\"$L5\",\"federated-learning-without-compromise\",{\"className\":\"blog-card\",\"href\":\"/blog/federated-learning-without-compromise\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"May 29, 2022\",\" \",\"·\",\" \",2,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Federated Learning Without Compromise\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Privacy-preserving machine learning that maintains model quality through novel aggregation protocols.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Federated Learning\",{\"className\":\"blog-tag\",\"children\":\"Federated Learning\"}],[\"$\",\"span\",\"Privacy\",{\"className\":\"blog-tag\",\"children\":\"Privacy\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"48:[\"$\",\"$L5\",\"federated-learning\",{\"className\":\"blog-card\",\"href\":\"/blog/federated-learning\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"May 8, 2022\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Federated Learning for Open AI\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How federated learning enables collaborative model training while preserving data privacy.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Training\",{\"className\":\"blog-tag\",\"children\":\"Training\"}],[\"$\",\"span\",\"Privacy\",{\"className\":\"blog-tag\",\"children\":\"Privacy\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"49:[\"$\",\"$L5\",\"experience-ledgers\",{\"className\":\"blog-card\",\"href\":\"/blog/experience-ledgers\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Feb 13, 2022\",\" \",\"·\",\" \",3,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Experience Ledgers: Persistent Memory for AI Agents\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Introducing experience ledgers, a framework for giving AI agents persistent, verifiable memory.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Agents\",{\"className\":\"blog-tag\",\"children\":\"Agents\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"4a:[\"$\",\"$L5\",\"case-for-decentralized-science\",{\"className\":\"blog-card\",\"href\":\"/blog/case-for-decentralized-science\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 7, 2021\",\" \",\"·\",\" \",2,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"The Case for Decentralized Science\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"A manifesto for decentralized science (DeSci) and its application to AI research.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"DeSci\",{\"className\":\"blog-tag\",\"children\":\"DeSci\"}],[\"$\",\"span\",\"Vision\",{\"className\":\"blog-tag\",\"children\":\"Vision\"}],[\"$\",\"span\",\"Governance\",{\"className\":\"blog-tag\",\"children\":\"Governance\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"4b:[\"$\",\"$L5\",\"case-for-desci\",{\"className\":\"blog-card\",\"href\":\"/blog/case-for-desci\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Nov 7, 2021\",\" \",\"·\",\" \",2,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"The Case for Decentralized Science\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"Why scientific research needs decentralization, and how blockchain can help.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"DeSci\",{\"className\":\"blog-tag\",\"children\":\"DeSci\"}],[\"$\",\"span\",\"Vision\",{\"className\":\"blog-tag\",\"children\":\"Vision\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"4c:[\"$\",\"$L5\",\"training-llms-collective-intelligence\",{\"className\":\"blog-card\",\"href\":\"/blog/training-llms-collective-intelligence\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Jul 21, 2021\",\" \",\"·\",\" \",2,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Training LLMs on Collective Intelligence\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"How we're approaching training data curation to capture humanity's collective intelligence.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"blog-tag\",\"children\":\"Research\"}],[\"$\",\"span\",\"Training\",{\"className\":\"blog-tag\",\"children\":\"Training\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"4d:[\"$\",\"$L5\",\"introducing-zen\",{\"className\":\"blog-card\",\"href\":\"/blog/introducing-zen\",\"data-upstream\":\"$undefined\",\"children\":[[\"$\",\"div\",null,{\"className\":\"blog-card-meta\",\"children\":[\"Mar 14, 2021\",\" \",\"·\",\" \",1,\" min read\"]}],[\"$\",\"h2\",null,{\"className\":\"blog-card-title\",\"children\":\"Introducing Zen: Open AI for the Open Web\"}],[\"$\",\"p\",null,{\"className\":\"blog-card-desc\",\"children\":\"We're launching Zen, an open research initiative to build AI that serves everyone.\"}],[\"$\",\"div\",null,{\"className\":\"blog-card-tags\",\"children\":[[\"$\",\"span\",\"Announcement\",{\"className\":\"blog-tag\",\"children\":\"Announcement\"}],[\"$\",\"span\",\"Vision\",{\"className\":\"blog-tag\",\"children\":\"Vision\"}]]}]]}]\n"])</script><script>self.__next_f.push([1,"53:[[\"$\",\"meta\",\"0\",{\"charSet\":\"utf-8\"}],[\"$\",\"meta\",\"1\",{\"name\":\"viewport\",\"content\":\"width=device-width, initial-scale=1\"}]]\n4f:null\n"])</script><script>self.__next_f.push([1,"51:{\"metadata\":[[\"$\",\"title\",\"0\",{\"children\":\"Zen Blog — research notes \u0026 releases\"}],[\"$\",\"meta\",\"1\",{\"name\":\"description\",\"content\":\"Architecture deep-dives, training research, and release notes from Zen LM, the open model family of Zoo Labs Foundation.\"}],[\"$\",\"meta\",\"2\",{\"name\":\"keywords\",\"content\":\"AI, LLM, Agentic AI, Code Generation, Zen Coder, Multimodal, Open Source, Machine Learning\"}]],\"error\":null,\"digest\":\"$undefined\"}\n"])</script><script>self.__next_f.push([1,"56:\"$51:metadata\"\n"])</script></body></html>