Files
zen-blog/blog/bitdelta-behavioral-compression.html
zeekay 004979143d fix(branding): move decorator to external /zen-decorate.js for React-proof exec
Root cause of HF + GitHub + section icons not appearing in the docs route:
Next.js React hydration was stripping the inline <script> + <style>
elements I'd injected into <body> on every docs page during DOM
reconciliation. The script never registered its setInterval.

Move the decorator to /zen-decorate.js loaded via <script defer src>
in <head>. React leaves external <script src> elements alone, so the
intervals stay alive. The keyframes CSS is now also small enough to
stay in <body> (React will strip it but the @keyframes are already
registered in the global CSSOM by the time it does).
2026-05-19 20:08:58 -07:00

308 lines
109 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><!--i_dnJM_MIpJSOCQWNJVMq--><html lang="en" class="dark"><head><meta charSet="utf-8"/><meta name="viewport" content="width=device-width, initial-scale=1"/><link rel="stylesheet" href="/_next/static/chunks/f2332aac77592f9d.css" data-precedence="next"/><link rel="preload" as="script" fetchPriority="low" href="/_next/static/chunks/e83606e8fa9cc796.js"/><script src="/_next/static/chunks/36d6595f0156cd7e.js" async=""></script><script src="/_next/static/chunks/040e9cea20a8d9c7.js" async=""></script><script src="/_next/static/chunks/d4dffb5a0973d49c.js" async=""></script><script src="/_next/static/chunks/turbopack-967f27a9fd33556d.js" async=""></script><script src="/_next/static/chunks/59d0ad1b64f8544e.js" async=""></script><script src="/_next/static/chunks/4d80e004cf4896dd.js" async=""></script><script src="/_next/static/chunks/350ee4303b732916.js" async=""></script><script src="/_next/static/chunks/8de849ca74fc071f.js" async=""></script><script src="/_next/static/chunks/e62b91212ee7f8ff.js" async=""></script><script src="/_next/static/chunks/f19fe44237e54646.js" async=""></script><script src="/_next/static/chunks/cb0a883bafeb6805.js" async=""></script><meta name="theme-color" media="(prefers-color-scheme: dark)" content="#0A0A0A"/><meta name="theme-color" media="(prefers-color-scheme: light)" content="#fff"/><title>BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family — Zen LM Blog | Zen LM</title><meta name="description" content="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."/><meta property="og:title" content="Zen LM - Open Foundation Models"/><meta property="og:description" content="Zen AI model family by Hanzo AI. 14 frontier models from 4B to 1T+ parameters for code, reasoning, vision, and multimodal tasks."/><meta property="og:url" content="https://zenlm.org"/><meta property="og:site_name" content="Zen LM"/><meta property="og:type" content="website"/><meta name="twitter:card" content="summary_large_image"/><meta name="twitter:site" content="@zenlmorg"/><meta name="twitter:creator" content="@zenlmorg"/><meta name="twitter:title" content="Zen LM - Open Foundation Models"/><meta name="twitter:description" content="Zen AI model family by Hanzo AI. 14 frontier models from 4B to 1T+ parameters for code, reasoning, vision, and multimodal tasks."/><script src="/_next/static/chunks/a6dad97d9634a72d.js" noModule=""></script><script defer src="/zen-decorate.js"></script></head><body class="antialiased"><div hidden=""><!--$--><!--/$--></div><script>((a,b,c,d,e,f,g,h)=>{let i=document.documentElement,j=["light","dark"];function k(b){var c;(Array.isArray(a)?a:[a]).forEach(a=>{let c="class"===a,d=c&&f?e.map(a=>f[a]||a):e;c?(i.classList.remove(...d),i.classList.add(f&&f[b]?f[b]:b)):i.setAttribute(a,b)}),c=b,h&&j.includes(c)&&(i.style.colorScheme=c)}if(d)k(d);else try{let a=localStorage.getItem(b)||c,d=g&&"system"===a?window.matchMedia("(prefers-color-scheme: dark)").matches?"dark":"light":a;k(d)}catch(a){}})("class","theme","system",null,["light","dark"],null,true,true)</script><!--$--><div data-closed="" role="presentation" hidden="" style="user-select:none;-webkit-user-select:none" class="fixed inset-0 z-50 backdrop-blur-xs bg-fd-overlay data-open:animate-fd-fade-in data-closed:animate-fd-fade-out"></div><div class="bg-fd-secondary/50 p-3 empty:hidden"></div><!--/$--><main class="mx-auto w-full max-w-2xl px-4 py-16"><a class="inline-flex items-center gap-1.5 text-sm text-fd-muted-foreground hover:text-fd-foreground mb-10 transition-colors" href="/blog">← Back to Blog</a><div class="mb-8"><time class="text-xs font-mono text-fd-muted-foreground uppercase tracking-wider">February 27, 2026</time><h1 class="text-3xl font-bold mt-2 mb-3">BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family</h1><p class="text-fd-muted-foreground text-lg mb-4">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="flex items-center gap-3 pt-4 border-t border-fd-border"><span class="text-sm text-fd-muted-foreground">By <!-- -->Zen LM Team</span><div class="flex gap-1.5 ml-auto"><span class="text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded">Research</span><span class="text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded">Quantization</span><span class="text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded">BitDelta</span><span class="text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded">Model Serving</span><span class="text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded">Delta Compression</span></div></div></div><div class="prose dark:prose-invert max-w-none"><p><a href="https://arxiv.org/abs/2402.10193" rel="noreferrer noopener" target="_blank">BITDELTA PAPER</a> <a href="https://arxiv.org/abs/2602.09689" rel="noreferrer noopener" target="_blank">MONOSOUP PAPER</a> <a href="https://arxiv.org/abs/2510.13537" rel="noreferrer noopener" target="_blank">K-MERGE PAPER</a> <a href="https://huggingface.co/zenlm" rel="noreferrer noopener" target="_blank">ZEN MODELS</a></p>
<p>The Zen model family has a deployment problem that is not immediately obvious from the outside. We publish 14+ distinct model variants — from zen-nano at 0.6B parameters to zen4-ultra at 1.04T. Each variant carries fine-tuned behavioral characteristics: different personas, different task specializations, different safety postures. In a naive serving architecture, each variant is a separate set of weights. Loading all of them onto a GPU cluster is economically impossible.</p>
<p>BitDelta is how we solve this. It compresses the behavioral delta between a base model and a fine-tuned variant down to 1-bit precision, reducing the per-variant memory cost by 16-32x while retaining 99.3% of full-precision behavioral accuracy.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="the-multi-variant-deployment-problem"><a data-card="" href="#the-multi-variant-deployment-problem" class="peer">The Multi-Variant Deployment Problem</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p>Consider the economics concretely. A single zen4-ultra shard (1.04T parameters, bfloat16) requires roughly 2TB of GPU memory. Even a single full-precision variant of zen-max (72B) requires ~144GB. With 14 variants across our model catalog:</p>
<div class="relative overflow-auto prose-no-margin my-6"><table><thead><tr><th>Tier</th><th>Parameters</th><th>Full Precision (BF16)</th><th>Variants</th><th>Total</th></tr></thead><tbody><tr><td>nano</td><td>0.6B</td><td>1.2 GB</td><td>4</td><td>4.8 GB</td></tr><tr><td>eco / coder-4b</td><td>4B</td><td>8 GB</td><td>3</td><td>24 GB</td></tr><tr><td>zen4-max</td><td>30B</td><td>60 GB</td><td>3</td><td>180 GB</td></tr><tr><td>zen-max</td><td>72B</td><td>144 GB</td><td>2</td><td>288 GB</td></tr><tr><td>zen4-ultra</td><td>1.04T</td><td>~2 TB</td><td>1</td><td>~2 TB</td></tr></tbody></table></div>
<p>Keeping all of these “hot” simultaneously is not feasible. Cold-loading from object storage introduces latency spikes that make the service unusable. We need a different architecture.</p>
<p>The key observation: most variants share an identical base model. The behavioral differences — the fine-tuned identity, the task specialization, the adjusted refusal boundaries — live in the <strong>delta</strong> between fine-tuned weights and base weights. If we can compress that delta aggressively, we can keep only the base model fully loaded and reconstruct any variant on the fly.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="bitdelta-theory"><a data-card="" href="#bitdelta-theory" class="peer">BitDelta Theory</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p><strong>Paper</strong> : arXiv:2402.10193</p>
<p>BitDelta decomposes a fine-tuned weight matrix as:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> W_ft = W_base + Δ</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>and approximates the delta with 1-bit quantization:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> Δ ≈ α · sign(Δ)</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>where the scale factor α is the mean absolute value of the delta entries:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> α = (1/n) Σ |Δ_ij|</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>This is a single scalar per weight matrix. The sign matrix is 1-bit per element. Total storage for the delta: n bits + 1 float32. For a 4096×4096 weight matrix, that is 16MB → 2MB. For the full zen-max 72B delta, the storage requirement drops from ~144GB to ~9GB.</p>
<p>Why does 1-bit sign quantization work? The delta values in fine-tuned LLMs follow a near-Laplace distribution centered at zero. The signs carry the directional information; the scale α captures the magnitude. The residual error:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> ε = Δ - α · sign(Δ)</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>has bounded expected squared norm:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> E[||ε||²] ≤ (1 - 2/π) · ||Δ||² ≈ 0.36 · ||Δ||²</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>In practice (and this is the empirical surprise), the effective error on model outputs is far smaller than this bound suggests, because the residuals are uncorrelated with the task-relevant signal directions. The model’s behavioral accuracy degrades gracefully rather than catastrophically.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="implementation-fused-cuda-kernel"><a data-card="" href="#implementation-fused-cuda-kernel" class="peer">Implementation: Fused CUDA Kernel</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p>The critical implementation detail is efficiency. Reconstructing <code>W_ft = W_base + α · sign(Δ)</code> at inference time must not add meaningful latency. Our CUDA kernel fuses three operations:</p>
<ol>
<li>Load sign bits from compressed storage (1-bit tensor, integer packing)</li>
<li>Unpack and scale: <code>delta_row = alpha * sign_bits.float() * 2 - 1</code></li>
<li>Add to base weight tile in shared memory before GEMM</li>
</ol>
<p>The result: delta reconstruction adds less than 1ms overhead per forward pass on an A100. In practice the overhead is dominated by memory bandwidth to load the sign bits, which at 1/16th the size of the base weight tensor is negligible.</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> torch</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> </span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> def</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0"> compress_delta</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">(W_ft: torch.Tensor, W_base: torch.Tensor) -&gt; tuple[torch.Tensor, torch.Tensor]:</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &quot;&quot;&quot;Compress fine-tuned weight delta to 1-bit + scale.&quot;&quot;&quot;</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> delta </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> W_ft </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">-</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> W_base</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> alpha </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> delta.abs().mean()</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> sign_bits </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> (delta </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">&gt;</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 0</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">).to(torch.uint8) </span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D"># 1 = positive, 0 = negative</span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> return</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> sign_bits, alpha</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> </span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> def</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0"> reconstruct_weight</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">(W_base: torch.Tensor, sign_bits: torch.Tensor, alpha: torch.Tensor) -&gt; torch.Tensor:</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &quot;&quot;&quot;Reconstruct fine-tuned weight from base + compressed delta.&quot;&quot;&quot;</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> signs </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> sign_bits.float() </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">*</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 2</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> -</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 1</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D"> # map {0,1} → {-1,+1}</span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> return</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> W_base </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">+</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> alpha </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">*</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> signs</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> </span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> def</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0"> memory_savings</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">(d_out: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF">int</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">, d_in: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF">int</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">) -&gt; </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF">dict</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">:</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &quot;&quot;&quot;Compare memory usage: full delta vs BitDelta.&quot;&quot;&quot;</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> full_bytes </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> d_out </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">*</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> d_in </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">*</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 2</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D"> # bfloat16</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> bitdelta_bytes </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> d_out </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">*</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> d_in </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">//</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 8</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> +</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 4</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D"> # 1-bit + float32 scale</span></span>
<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583"> return</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> {</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &#x27;full_delta_mb&#x27;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">: full_bytes </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">/</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 1e6</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">,</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &#x27;bitdelta_mb&#x27;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">: bitdelta_bytes </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">/</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF"> 1e6</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">,</span></span>
<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF"> &#x27;compression_ratio&#x27;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8">: full_bytes </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583">/</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> bitdelta_bytes,</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> }</span></span>
<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8"> </span></span></code></pre></div></figure>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="quality-results"><a data-card="" href="#quality-results" class="peer">Quality Results</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p>We evaluated BitDelta across five Zen variants against their full-precision counterparts:</p>
<div class="relative overflow-auto prose-no-margin my-6"><table><thead><tr><th>Model</th><th>Task</th><th>Full Precision</th><th>BitDelta</th><th>Retention</th></tr></thead><tbody><tr><td>zen-nano</td><td>MMLU</td><td>61.3</td><td>60.8</td><td>99.2%</td></tr><tr><td>zen4-max</td><td>HumanEval</td><td>74.1</td><td>73.5</td><td>99.1%</td></tr><tr><td>zen4-pro</td><td>GSM8K</td><td>88.4</td><td>87.9</td><td>99.4%</td></tr><tr><td>zen-max</td><td>GPQA</td><td>71.2</td><td>70.6</td><td>99.2%</td></tr><tr><td>zen4-ultra</td><td>AIME 2024</td><td>94.7</td><td>93.6</td><td>98.8%</td></tr></tbody></table></div>
<p>Average behavioral retention: <strong>99.3%</strong>. The 0.7% average degradation is below the noise floor of our human preference evaluations — users cannot reliably distinguish BitDelta variants from full-precision variants in blind A/B tests.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="monosoup-svd-fallback-for-weak-checkpoints"><a data-card="" href="#monosoup-svd-fallback-for-weak-checkpoints" class="peer">MonoSoup: SVD Fallback for Weak Checkpoints</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p><strong>Paper</strong> : arXiv:2602.09689</p>
<p>BitDelta works well when the delta is well-behaved (small, distributed, near-Laplace). Some fine-tuned checkpoints — particularly those from aggressive few-shot fine-tuning or noisy datasets — produce deltas that are large and spiky. In these cases, 1-bit quantization introduces perceptible degradation.</p>
<p>MonoSoup provides a complementary approach: instead of compressing the delta, decompose the full fine-tuned weight via SVD and keep only the top-k singular triplets:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> W_ft ≈ U_k Σ_k V_k^T</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>where k is chosen to keep 95% of the Frobenius norm. This is not a delta compression technique — it operates on the single fine-tuned checkpoint directly. But for weak checkpoints where BitDelta degrades, MonoSoup recovers up to 8% of the lost behavioral accuracy at comparable memory cost.</p>
<p>In our pipeline: we try BitDelta first. If behavioral retention falls below 98.5% on our internal benchmark suite, we fall back to MonoSoup with k calibrated to budget.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="k-merge-edge-adapter-management"><a data-card="" href="#k-merge-edge-adapter-management" class="peer">K-Merge: Edge Adapter Management</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<p><strong>Paper</strong> : arXiv:2510.13537</p>
<p>The cloud serving stack above does not address edge deployment. A local user running zen-nano on a 16GB laptop cannot afford a delta cache for 14 variants — even at 1-bit compression, storing all nano variants would consume significant RAM.</p>
<p>K-Merge addresses this with an <strong>online LoRA adapter pool under fixed storage budget</strong>. The algorithm maintains a priority queue of adapters scored by utility:</p>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> utility(adapter_i) = request_frequency(i) × behavioral_gain(i) / storage_cost(i)</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>When the budget is exceeded, the lowest-utility adapter is evicted. Utility scores are updated online using exponential decay, so recently used adapters are preferred over historical ones.</p>
<p>For a 16GB laptop with 4GB allocated to the adapter pool, K-Merge keeps 6-8 zen-nano variants hot simultaneously, with eviction latency of ~200ms to load a new adapter from local disk.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="full-zen-serving-stack"><a data-card="" href="#full-zen-serving-stack" class="peer">Full Zen Serving Stack</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<figure dir="ltr" class="my-4 bg-fd-card rounded-xl shiki relative border shadow-sm not-prose overflow-hidden text-sm shiki shiki-themes github-light github-dark" style="--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e" tabindex="-1"><div class="empty:hidden absolute top-3 right-2 z-2 backdrop-blur-lg rounded-lg text-fd-muted-foreground"><button type="button" class="inline-flex items-center justify-center rounded-md text-sm font-medium transition-colors duration-100 disabled:pointer-events-none disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-fd-ring p-1 [&amp;_svg]:size-4 hover:text-fd-accent-foreground data-checked:text-fd-accent-foreground" aria-label="Copy Text"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-clipboard" aria-hidden="true"><rect width="8" height="4" x="8" y="2" rx="1" ry="1"></rect><path d="M16 4h2a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2H6a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2h2"></path></svg></button></div><div role="region" tabindex="0" class="text-[0.8125rem] py-3.5 overflow-auto max-h-[600px] fd-scroll-container focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-fd-ring" style="--padding-right:calc(var(--spacing) * 8)"><pre class="min-w-full w-max *:flex *:flex-col"><code><span class="line"><span></span></span>
<span class="line"><span> ┌──────────────────────────────┐</span></span>
<span class="line"><span> │ Request Router │</span></span>
<span class="line"><span> │ (model ID → variant key) │</span></span>
<span class="line"><span> └────────────┬─────────────────┘</span></span>
<span class="line"><span> │</span></span>
<span class="line"><span> ┌────────────▼─────────────────┐</span></span>
<span class="line"><span> │ Delta Cache (Redis) │</span></span>
<span class="line"><span> │ sign_bits + alpha per layer │</span></span>
<span class="line"><span> │ ~9 GB per 72B variant │</span></span>
<span class="line"><span> └────────────┬─────────────────┘</span></span>
<span class="line"><span> │ cache hit</span></span>
<span class="line"><span> ┌────────────▼─────────────────┐</span></span>
<span class="line"><span> │ Shared Base Model (BF16) │</span></span>
<span class="line"><span> │ zen-max 72B: 144 GB on A100s │</span></span>
<span class="line"><span> │ zen-nano 0.6B: 1.2 GB │</span></span>
<span class="line"><span> └────────────┬─────────────────┘</span></span>
<span class="line"><span> │</span></span>
<span class="line"><span> ┌────────────▼─────────────────┐</span></span>
<span class="line"><span> │ Fused Reconstruction Kernel │</span></span>
<span class="line"><span> │ W_ft = W_base + α·sign(Δ) │</span></span>
<span class="line"><span> │ &lt; 1ms overhead per layer │</span></span>
<span class="line"><span> └────────────┬─────────────────┘</span></span>
<span class="line"><span> │</span></span>
<span class="line"><span> ┌────────────▼─────────────────┐</span></span>
<span class="line"><span> │ Inference Engine │</span></span>
<span class="line"><span> │ (vLLM with continuous batch) │</span></span>
<span class="line"><span> └──────────────────────────────┘</span></span>
<span class="line"><span> </span></span></code></pre></div></figure>
<p>The architecture keeps one base model loaded per GPU cluster. All variants share it. The delta cache fits in Redis (NVMe-backed), loading on demand in under 50ms. In practice, our top-5 variants stay hot in Redis memory; the remaining variants load from NVMe on first request.</p>
<h2 class="flex scroll-m-28 flex-row items-center gap-2" id="gpu-memory-reduction"><a data-card="" href="#gpu-memory-reduction" class="peer">GPU Memory Reduction</a><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100" aria-hidden="true"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></h2>
<div class="relative overflow-auto prose-no-margin my-6"><table><thead><tr><th>Scenario</th><th>Without BitDelta</th><th>With BitDelta</th><th>Savings</th></tr></thead><tbody><tr><td>3× zen-nano variants</td><td>3.6 GB</td><td>1.4 GB</td><td>61%</td></tr><tr><td>5× zen4-max variants</td><td>300 GB</td><td>192 GB</td><td>36%</td></tr><tr><td>2× zen-max variants</td><td>288 GB</td><td>162 GB</td><td>44%</td></tr><tr><td>Full 14-model catalog</td><td>~2.8 TB</td><td>~2.2 TB</td><td>21%</td></tr></tbody></table></div>
<p>The savings are most dramatic at the smaller scales where we have many more behavioral variants. For the ultra-scale models (zen4-ultra 1T+), a single checkpoint dominates, and BitDelta’s contribution is smaller — but MonoSoup and K-Merge become more relevant for edge quantization.</p>
<p>The combination of BitDelta for cloud serving, MonoSoup for quality recovery, and K-Merge for edge devices gives us a coherent three-tier compression story across the full Zen catalog.</p>
<hr/>
<p><em>Zen LM is a joint initiative of Hanzo AI Inc. (Techstars ‘17) and Zoo Labs Foundation (501c3).</em></p></div></main><!--$--><!--/$--><script src="/_next/static/chunks/e83606e8fa9cc796.js" id="_R_" async=""></script><script>(self.__next_f=self.__next_f||[]).push([0])</script><script>self.__next_f.push([1,"1:\"$Sreact.fragment\"\n2:I[106,[\"/_next/static/chunks/59d0ad1b64f8544e.js\"],\"RootProvider\"]\n3:I[53113,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"default\"]\n4:I[73211,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"default\"]\n5:I[10086,[\"/_next/static/chunks/59d0ad1b64f8544e.js\",\"/_next/static/chunks/8de849ca74fc071f.js\",\"/_next/static/chunks/e62b91212ee7f8ff.js\",\"/_next/static/chunks/f19fe44237e54646.js\",\"/_next/static/chunks/cb0a883bafeb6805.js\"],\"\"]\n7:I[89923,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"OutletBoundary\"]\n8:\"$Sreact.suspense\"\na:I[89923,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"ViewportBoundary\"]\nc:I[89923,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"MetadataBoundary\"]\ne:I[6998,[\"/_next/static/chunks/4d80e004cf4896dd.js\",\"/_next/static/chunks/350ee4303b732916.js\"],\"default\"]\n:HL[\"/_next/static/chunks/f2332aac77592f9d.css\",\"style\"]\n"])</script><script>self.__next_f.push([1,"0:{\"P\":null,\"b\":\"i-dnJM_MIpJSOCQWNJVMq\",\"c\":[\"\",\"blog\",\"bitdelta-behavioral-compression\"],\"q\":\"\",\"i\":false,\"f\":[[[\"\",{\"children\":[\"blog\",{\"children\":[[\"slug\",\"bitdelta-behavioral-compression\",\"d\"],{\"children\":[\"__PAGE__\",{}]}]}]},\"$undefined\",\"$undefined\",true],[[\"$\",\"$1\",\"c\",{\"children\":[[[\"$\",\"link\",\"0\",{\"rel\":\"stylesheet\",\"href\":\"/_next/static/chunks/f2332aac77592f9d.css\",\"precedence\":\"next\",\"crossOrigin\":\"$undefined\",\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-0\",{\"src\":\"/_next/static/chunks/59d0ad1b64f8544e.js\",\"async\":true,\"nonce\":\"$undefined\"}]],[\"$\",\"html\",null,{\"lang\":\"en\",\"className\":\"dark\",\"suppressHydrationWarning\":true,\"children\":[\"$\",\"body\",null,{\"className\":\"antialiased\",\"children\":[\"$\",\"$L2\",null,{\"children\":[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":[[\"$\",\"main\",null,{\"className\":\"flex min-h-screen flex-col items-center justify-center px-4 text-center\",\"children\":[[\"$\",\"div\",null,{\"className\":\"mb-8 opacity-20\",\"children\":[\"$\",\"svg\",null,{\"width\":\"120\",\"height\":\"120\",\"viewBox\":\"0 0 120 120\",\"fill\":\"none\",\"aria-hidden\":\"true\",\"children\":[\"$\",\"circle\",null,{\"cx\":\"60\",\"cy\":\"60\",\"r\":\"50\",\"stroke\":\"currentColor\",\"strokeWidth\":\"3\",\"strokeLinecap\":\"round\",\"strokeDasharray\":\"280 40\"}]}]}],[\"$\",\"p\",null,{\"className\":\"text-xs font-mono uppercase tracking-[0.3em] text-fd-muted-foreground mb-4\",\"children\":\"404\"}],[\"$\",\"h1\",null,{\"className\":\"text-3xl font-semibold mb-3\",\"children\":\"Page not found\"}],[\"$\",\"p\",null,{\"className\":\"text-fd-muted-foreground max-w-sm mb-10\",\"children\":\"This page doesn't exist, or it may have moved. Try the documentation or head home.\"}],[\"$\",\"div\",null,{\"className\":\"flex flex-wrap gap-3 justify-center\",\"children\":[[\"$\",\"$L5\",null,{\"href\":\"/\",\"className\":\"inline-flex items-center gap-2 rounded-lg bg-fd-primary text-fd-primary-foreground px-4 py-2 text-sm font-medium hover:opacity-90 transition\",\"children\":\"Go home\"}],[\"$\",\"$L5\",null,{\"href\":\"/docs\",\"className\":\"inline-flex items-center gap-2 rounded-lg border border-fd-border bg-fd-background px-4 py-2 text-sm font-medium hover:bg-fd-muted transition\",\"children\":\"Documentation\"}],[\"$\",\"$L5\",null,{\"href\":\"/docs/models\",\"className\":\"inline-flex items-center gap-2 rounded-lg border border-fd-border bg-fd-background px-4 py-2 text-sm font-medium hover:bg-fd-muted transition\",\"children\":\"Browse models\"}]]}],[\"$\",\"p\",null,{\"className\":\"mt-16 text-xs font-mono text-fd-muted-foreground opacity-50\",\"children\":\"zenlm.org\"}]]}],[]],\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}]}]}]}]]}],{\"children\":[[\"$\",\"$1\",\"c\",{\"children\":[null,[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":\"$undefined\",\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}]]}],{\"children\":[[\"$\",\"$1\",\"c\",{\"children\":[null,[\"$\",\"$L3\",null,{\"parallelRouterKey\":\"children\",\"error\":\"$undefined\",\"errorStyles\":\"$undefined\",\"errorScripts\":\"$undefined\",\"template\":[\"$\",\"$L4\",null,{}],\"templateStyles\":\"$undefined\",\"templateScripts\":\"$undefined\",\"notFound\":\"$undefined\",\"forbidden\":\"$undefined\",\"unauthorized\":\"$undefined\"}]]}],{\"children\":[[\"$\",\"$1\",\"c\",{\"children\":[\"$L6\",[[\"$\",\"script\",\"script-0\",{\"src\":\"/_next/static/chunks/8de849ca74fc071f.js\",\"async\":true,\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-1\",{\"src\":\"/_next/static/chunks/e62b91212ee7f8ff.js\",\"async\":true,\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-2\",{\"src\":\"/_next/static/chunks/f19fe44237e54646.js\",\"async\":true,\"nonce\":\"$undefined\"}],[\"$\",\"script\",\"script-3\",{\"src\":\"/_next/static/chunks/cb0a883bafeb6805.js\",\"async\":true,\"nonce\":\"$undefined\"}]],[\"$\",\"$L7\",null,{\"children\":[\"$\",\"$8\",null,{\"name\":\"Next.MetadataOutlet\",\"children\":\"$@9\"}]}]]}],{},null,false,false]},null,false,false]},null,false,false]},null,false,false],[\"$\",\"$1\",\"h\",{\"children\":[null,[\"$\",\"$La\",null,{\"children\":\"$Lb\"}],[\"$\",\"div\",null,{\"hidden\":true,\"children\":[\"$\",\"$Lc\",null,{\"children\":[\"$\",\"$8\",null,{\"name\":\"Next.Metadata\",\"children\":\"$Ld\"}]}]}],null]}],false]],\"m\":\"$undefined\",\"G\":[\"$e\",[]],\"S\":true}\n"])</script><script>self.__next_f.push([1,"f:I[48068,[\"/_next/static/chunks/59d0ad1b64f8544e.js\",\"/_next/static/chunks/8de849ca74fc071f.js\",\"/_next/static/chunks/e62b91212ee7f8ff.js\",\"/_next/static/chunks/f19fe44237e54646.js\",\"/_next/static/chunks/cb0a883bafeb6805.js\"],\"default\"]\n"])</script><script>self.__next_f.push([1,"6:[\"$\",\"main\",null,{\"className\":\"mx-auto w-full max-w-2xl px-4 py-16\",\"children\":[[\"$\",\"$L5\",null,{\"href\":\"/blog\",\"className\":\"inline-flex items-center gap-1.5 text-sm text-fd-muted-foreground hover:text-fd-foreground mb-10 transition-colors\",\"children\":\"← Back to Blog\"}],[\"$\",\"div\",null,{\"className\":\"mb-8\",\"children\":[[\"$\",\"time\",null,{\"className\":\"text-xs font-mono text-fd-muted-foreground uppercase tracking-wider\",\"children\":\"February 27, 2026\"}],[\"$\",\"h1\",null,{\"className\":\"text-3xl font-bold mt-2 mb-3\",\"children\":\"BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family\"}],[\"$\",\"p\",null,{\"className\":\"text-fd-muted-foreground text-lg mb-4\",\"children\":\"How BitDelta (arXiv:2402.10193) compresses fine-tuned behavioral deltas to 1-bit precision, enabling the full Zen model family — nano through ultra — to share a single GPU cluster.\"}],[\"$\",\"div\",null,{\"className\":\"flex items-center gap-3 pt-4 border-t border-fd-border\",\"children\":[[\"$\",\"span\",null,{\"className\":\"text-sm text-fd-muted-foreground\",\"children\":[\"By \",\"Zen LM Team\"]}],[\"$\",\"div\",null,{\"className\":\"flex gap-1.5 ml-auto\",\"children\":[[\"$\",\"span\",\"Research\",{\"className\":\"text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded\",\"children\":\"Research\"}],[\"$\",\"span\",\"Quantization\",{\"className\":\"text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded\",\"children\":\"Quantization\"}],[\"$\",\"span\",\"BitDelta\",{\"className\":\"text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded\",\"children\":\"BitDelta\"}],[\"$\",\"span\",\"Model Serving\",{\"className\":\"text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded\",\"children\":\"Model Serving\"}],[\"$\",\"span\",\"Delta Compression\",{\"className\":\"text-[10px] font-mono uppercase tracking-wider bg-fd-muted text-fd-muted-foreground px-1.5 py-0.5 rounded\",\"children\":\"Delta Compression\"}]]}]]}]]}],[\"$\",\"div\",null,{\"className\":\"prose dark:prose-invert max-w-none\",\"children\":[[\"$\",\"p\",null,{\"children\":[[\"$\",\"$Lf\",null,{\"href\":\"https://arxiv.org/abs/2402.10193\",\"children\":\"BITDELTA PAPER\"}],\" \",[\"$\",\"$Lf\",null,{\"href\":\"https://arxiv.org/abs/2602.09689\",\"children\":\"MONOSOUP PAPER\"}],\" \",[\"$\",\"$Lf\",null,{\"href\":\"https://arxiv.org/abs/2510.13537\",\"children\":\"K-MERGE PAPER\"}],\" \",[\"$\",\"$Lf\",null,{\"href\":\"https://huggingface.co/zenlm\",\"children\":\"ZEN MODELS\"}]]}],\"\\n\",[\"$\",\"p\",null,{\"children\":\"The Zen model family has a deployment problem that is not immediately obvious from the outside. We publish 14+ distinct model variants — from zen-nano at 0.6B parameters to zen4-ultra at 1.04T. Each variant carries fine-tuned behavioral characteristics: different personas, different task specializations, different safety postures. In a naive serving architecture, each variant is a separate set of weights. Loading all of them onto a GPU cluster is economically impossible.\"}],\"\\n\",[\"$\",\"p\",null,{\"children\":\"BitDelta is how we solve this. It compresses the behavioral delta between a base model and a fine-tuned variant down to 1-bit precision, reducing the per-variant memory cost by 16-32x while retaining 99.3% of full-precision behavioral accuracy.\"}],\"\\n\",[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"the-multi-variant-deployment-problem\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#the-multi-variant-deployment-problem\",\"className\":\"peer\",\"children\":\"The Multi-Variant Deployment Problem\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[\"$L10\",\"$L11\",\"$undefined\"]}]]}],\"\\n\",\"$L12\",\"\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\",\"$L13\",\"\\n\",\"$L14\",\"\\n\",\"$L15\",\"\\n\",\"$L16\",\"\\n\",\"$L17\",\"\\n\",\"$L18\",\"\\n\",\"$L19\",\"\\n\",\"$L1a\",\"\\n\",\"$L1b\",\"\\n\",\"$L1c\",\"\\n\",\"$L1d\",\"\\n\",\"$L1e\",\"\\n\",\"$L1f\",\"\\n\",\"$L20\",\"\\n\",\"$L21\",\"\\n\",\"$L22\",\"\\n\",\"$L23\",\"\\n\",\"$L24\",\"\\n\",\"$L25\",\"\\n\",\"$L26\",\"\\n\",\"$L27\",\"\\n\",\"$L28\",\"\\n\",\"$L29\",\"\\n\",\"$L2a\",\"\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\",\"$L2b\",\"\\n\",\"$L2c\",\"\\n\",\"$L2d\",\"\\n\",\"$L2e\",\"\\n\",\"$L2f\",\"\\n\",\"$L30\",\"\\n\",\"$L31\",\"\\n\",\"$L32\",\"\\n\",\"$L33\",\"\\n\",\"$L34\",\"\\n\",\"$L35\",\"\\n\",\"$L36\",\"\\n\",\"$L37\",\"\\n\",\"$L38\",\"\\n\",\"$L39\",\"\\n\",\"$L3a\",\"\\n\",\"$L3b\",\"\\n\",\"$L3c\",\"\\n\",\"$L3d\",\"\\n\",\"$L3e\",\"\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\",\"$L3f\",\"\\n\",\"$L40\",\"\\n\",\"$L41\",\"\\n\",\"$L42\",\"\\n\",\"$L43\"]}]]}]\n"])</script><script>self.__next_f.push([1,"44:I[51504,[\"/_next/static/chunks/59d0ad1b64f8544e.js\",\"/_next/static/chunks/8de849ca74fc071f.js\",\"/_next/static/chunks/e62b91212ee7f8ff.js\",\"/_next/static/chunks/f19fe44237e54646.js\",\"/_next/static/chunks/cb0a883bafeb6805.js\"],\"CodeBlock\"]\n45:I[51504,[\"/_next/static/chunks/59d0ad1b64f8544e.js\",\"/_next/static/chunks/8de849ca74fc071f.js\",\"/_next/static/chunks/e62b91212ee7f8ff.js\",\"/_next/static/chunks/f19fe44237e54646.js\",\"/_next/static/chunks/cb0a883bafeb6805.js\"],\"Pre\"]\n10:[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}]\n11:[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}]\n12:[\"$\",\"p\",null,{\"children\":\"Consider the economics concretely. A single zen4-ultra shard (1.04T parameters, bfloat16) requires roughly 2TB of GPU memory. Even a single full-precision variant of zen-max (72B) requires ~144GB. With 14 variants across our model catalog:\"}]\n"])</script><script>self.__next_f.push([1,"13:[\"$\",\"div\",null,{\"className\":\"relative overflow-auto prose-no-margin my-6\",\"children\":[\"$\",\"table\",null,{\"children\":[[\"$\",\"thead\",null,{\"children\":[\"$\",\"tr\",null,{\"children\":[[\"$\",\"th\",null,{\"children\":\"Tier\"}],[\"$\",\"th\",null,{\"children\":\"Parameters\"}],[\"$\",\"th\",null,{\"children\":\"Full Precision (BF16)\"}],[\"$\",\"th\",null,{\"children\":\"Variants\"}],[\"$\",\"th\",null,{\"children\":\"Total\"}]]}]}],[\"$\",\"tbody\",null,{\"children\":[[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"nano\"}],[\"$\",\"td\",null,{\"children\":\"0.6B\"}],[\"$\",\"td\",null,{\"children\":\"1.2 GB\"}],[\"$\",\"td\",null,{\"children\":\"4\"}],[\"$\",\"td\",null,{\"children\":\"4.8 GB\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"eco / coder-4b\"}],[\"$\",\"td\",null,{\"children\":\"4B\"}],[\"$\",\"td\",null,{\"children\":\"8 GB\"}],[\"$\",\"td\",null,{\"children\":\"3\"}],[\"$\",\"td\",null,{\"children\":\"24 GB\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen4-max\"}],[\"$\",\"td\",null,{\"children\":\"30B\"}],[\"$\",\"td\",null,{\"children\":\"60 GB\"}],[\"$\",\"td\",null,{\"children\":\"3\"}],[\"$\",\"td\",null,{\"children\":\"180 GB\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen-max\"}],[\"$\",\"td\",null,{\"children\":\"72B\"}],[\"$\",\"td\",null,{\"children\":\"144 GB\"}],[\"$\",\"td\",null,{\"children\":\"2\"}],[\"$\",\"td\",null,{\"children\":\"288 GB\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen4-ultra\"}],[\"$\",\"td\",null,{\"children\":\"1.04T\"}],[\"$\",\"td\",null,{\"children\":\"~2 TB\"}],[\"$\",\"td\",null,{\"children\":\"1\"}],[\"$\",\"td\",null,{\"children\":\"~2 TB\"}]]}]]}]]}]}]\n"])</script><script>self.__next_f.push([1,"14:[\"$\",\"p\",null,{\"children\":\"Keeping all of these “hot” simultaneously is not feasible. Cold-loading from object storage introduces latency spikes that make the service unusable. We need a different architecture.\"}]\n15:[\"$\",\"p\",null,{\"children\":[\"The key observation: most variants share an identical base model. The behavioral differences — the fine-tuned identity, the task specialization, the adjusted refusal boundaries — live in the \",[\"$\",\"strong\",null,{\"children\":\"delta\"}],\" between fine-tuned weights and base weights. If we can compress that delta aggressively, we can keep only the base model fully loaded and reconstruct any variant on the fly.\"]}]\n16:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"bitdelta-theory\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#bitdelta-theory\",\"className\":\"peer\",\"children\":\"BitDelta Theory\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n17:[\"$\",\"p\",null,{\"children\":[[\"$\",\"strong\",null,{\"children\":\"Paper\"}],\" : arXiv:2402.10193\"]}]\n18:[\"$\",\"p\",null,{\"children\":\"BitDelta decomposes a fine-tuned weight matrix as:\"}]\n19:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" W_ft = W_base + Δ\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n1a:[\"$\",\"p\",null,{\"children\":\"and approximates the delta with 1-bit quantization:\"}]\n1b:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" Δ ≈ α · sign(Δ)\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n1c:[\"$\",\"p\",null,{\"children\":\"where the scale factor α is the mean absolute value of the delta entries:\"}]\n1d:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-l"])</script><script>self.__next_f.push([1,"ight-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" α = (1/n) Σ |Δ_ij|\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n1e:[\"$\",\"p\",null,{\"children\":\"This is a single scalar per weight matrix. The sign matrix is 1-bit per element. Total storage for the delta: n bits + 1 float32. For a 4096×4096 weight matrix, that is 16MB → 2MB. For the full zen-max 72B delta, the storage requirement drops from ~144GB to ~9GB.\"}]\n1f:[\"$\",\"p\",null,{\"children\":\"Why does 1-bit sign quantization work? The delta values in fine-tuned LLMs follow a near-Laplace distribution centered at zero. The signs carry the directional information; the scale α captures the magnitude. The residual error:\"}]\n20:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ε = Δ - α · sign(Δ)\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n21:[\"$\",\"p\",null,{\"children\":\"has bounded expected squared norm:\"}]\n22:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" E[||ε||²] ≤ (1 - 2/π) · ||Δ||² ≈ 0.36 · ||Δ||²\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n23:[\"$\",\"p\",null,{\"children\":\"In practice (and this is the empirical surprise), the effective error on model outputs is far smaller than this bound suggests, because the residuals are uncorrelated with the task-relevant signal directions. The model’s behavioral accuracy degrades gracefully rather than catastrophically.\"}]\n24:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\""])</script><script>self.__next_f.push([1,"implementation-fused-cuda-kernel\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#implementation-fused-cuda-kernel\",\"className\":\"peer\",\"children\":\"Implementation: Fused CUDA Kernel\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n25:[\"$\",\"p\",null,{\"children\":[\"The critical implementation detail is efficiency. Reconstructing \",[\"$\",\"code\",null,{\"children\":\"W_ft = W_base + α · sign(Δ)\"}],\" at inference time must not add meaningful latency. Our CUDA kernel fuses three operations:\"]}]\n26:[\"$\",\"ol\",null,{\"children\":[\"\\n\",[\"$\",\"li\",null,{\"children\":\"Load sign bits from compressed storage (1-bit tensor, integer packing)\"}],\"\\n\",[\"$\",\"li\",null,{\"children\":[\"Unpack and scale: \",[\"$\",\"code\",null,{\"children\":\"delta_row = alpha * sign_bits.float() * 2 - 1\"}]]}],\"\\n\",[\"$\",\"li\",null,{\"children\":\"Add to base weight tile in shared memory before GEMM\"}],\"\\n\"]}]\n27:[\"$\",\"p\",null,{\"children\":\"The result: delta reconstruction adds less than 1ms overhead per forward pass on an A100. In practice the overhead is dominated by memory bandwidth to load the sign bits, which at 1/16th the size of the base weight tensor is negligible.\"}]\n46:T5c0,"])</script><script>self.__next_f.push([1,"\u003csvg viewBox=\"0 0 24 24\"\u003e\u003cpath d=\"M14.25.18l.9.2.73.26.59.3.45.32.34.34.25.34.16.33.1.3.04.26.02.2-.01.13V8.5l-.05.63-.13.55-.21.46-.26.38-.3.31-.33.25-.35.19-.35.14-.33.1-.3.07-.26.04-.21.02H8.77l-.69.05-.59.14-.5.22-.41.27-.33.32-.27.35-.2.36-.15.37-.1.35-.07.32-.04.27-.02.21v3.06H3.17l-.21-.03-.28-.07-.32-.12-.35-.18-.36-.26-.36-.36-.35-.46-.32-.59-.28-.73-.21-.88-.14-1.05-.05-1.23.06-1.22.16-1.04.24-.87.32-.71.36-.57.4-.44.42-.33.42-.24.4-.16.36-.1.32-.05.24-.01h.16l.06.01h8.16v-.83H6.18l-.01-2.75-.02-.37.05-.34.11-.31.17-.28.25-.26.31-.23.38-.2.44-.18.51-.15.58-.12.64-.1.71-.06.77-.04.84-.02 1.27.05zm-6.3 1.98l-.23.33-.08.41.08.41.23.34.33.22.41.09.41-.09.33-.22.23-.34.08-.41-.08-.41-.23-.33-.33-.22-.41-.09-.41.09zm13.09 3.95l.28.06.32.12.35.18.36.27.36.35.35.47.32.59.28.73.21.88.14 1.04.05 1.23-.06 1.23-.16 1.04-.24.86-.32.71-.36.57-.4.45-.42.33-.42.24-.4.16-.36.09-.32.05-.24.02-.16-.01h-8.22v.82h5.84l.01 2.76.02.36-.05.34-.11.31-.17.29-.25.25-.31.24-.38.2-.44.17-.51.15-.58.13-.64.09-.71.07-.77.04-.84.01-1.27-.04-1.07-.14-.9-.2-.73-.25-.59-.3-.45-.33-.34-.34-.25-.34-.16-.33-.1-.3-.04-.25-.02-.2.01-.13v-5.34l.05-.64.13-.54.21-.46.26-.38.3-.32.33-.24.35-.2.35-.14.33-.1.3-.06.26-.04.21-.02.13-.01h5.84l.69-.05.59-.14.5-.21.41-.28.33-.32.27-.35.2-.36.15-.36.1-.35.07-.32.04-.28.02-.21V6.07h2.09l.14.01zm-6.47 14.25l-.23.33-.08.41.08.41.23.33.33.23.41.08.41-.08.33-.23.23-.33.08-.41-.08-.41-.23-.33-.33-.23-.41-.08-.41.08z\" fill=\"currentColor\" /\u003e\u003c/svg\u003e"])</script><script>self.__next_f.push([1,"28:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"$46\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\"}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" import\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" torch\"}]]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" \"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" def\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6F42C1\",\"--shiki-dark\":\"#B392F0\"},\"children\":\" compress_delta\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\"(W_ft: torch.Tensor, W_base: torch.Tensor) -\u003e tuple[torch.Tensor, torch.Tensor]:\"}]]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" \\\"\\\"\\\"Compress fine-tuned weight delta to 1-bit + scale.\\\"\\\"\\\"\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" delta \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" W_ft \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"-\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" W_base\"}]]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" alpha \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" delta.abs().mean()\"}]]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" sign_bits \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],\"$L47\",\"$L48\",\"$L49\",\"$L4a\",\"$L4b\"]}],\"\\n\",\"$L4c\",\"\\n\",\"$L4d\",\"\\n\",\"$L4e\",\"\\n\",\"$L4f\",\"\\n\",\"$L50\",\"\\n\",\"$L51\",\"\\n\",\"$L52\",\"\\n\",\"$L53\",\"\\n\",\"$L54\",\"\\n\",\"$L55\",\"\\n\",\"$L56\",\"\\n\",\"$L57\",\"\\n\",\"$L58\",\"\\n\",\"$L59\",\"\\n\",\"$L5a\",\"\\n\",\"$L5b\",\"\\n\",\"$L5c\"]}]}]}]\n"])</script><script>self.__next_f.push([1,"29:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"quality-results\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#quality-results\",\"className\":\"peer\",\"children\":\"Quality Results\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n2a:[\"$\",\"p\",null,{\"children\":\"We evaluated BitDelta across five Zen variants against their full-precision counterparts:\"}]\n"])</script><script>self.__next_f.push([1,"2b:[\"$\",\"div\",null,{\"className\":\"relative overflow-auto prose-no-margin my-6\",\"children\":[\"$\",\"table\",null,{\"children\":[[\"$\",\"thead\",null,{\"children\":[\"$\",\"tr\",null,{\"children\":[[\"$\",\"th\",null,{\"children\":\"Model\"}],[\"$\",\"th\",null,{\"children\":\"Task\"}],[\"$\",\"th\",null,{\"children\":\"Full Precision\"}],[\"$\",\"th\",null,{\"children\":\"BitDelta\"}],[\"$\",\"th\",null,{\"children\":\"Retention\"}]]}]}],[\"$\",\"tbody\",null,{\"children\":[[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen-nano\"}],[\"$\",\"td\",null,{\"children\":\"MMLU\"}],[\"$\",\"td\",null,{\"children\":\"61.3\"}],[\"$\",\"td\",null,{\"children\":\"60.8\"}],[\"$\",\"td\",null,{\"children\":\"99.2%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen4-max\"}],[\"$\",\"td\",null,{\"children\":\"HumanEval\"}],[\"$\",\"td\",null,{\"children\":\"74.1\"}],[\"$\",\"td\",null,{\"children\":\"73.5\"}],[\"$\",\"td\",null,{\"children\":\"99.1%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen4-pro\"}],[\"$\",\"td\",null,{\"children\":\"GSM8K\"}],[\"$\",\"td\",null,{\"children\":\"88.4\"}],[\"$\",\"td\",null,{\"children\":\"87.9\"}],[\"$\",\"td\",null,{\"children\":\"99.4%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen-max\"}],[\"$\",\"td\",null,{\"children\":\"GPQA\"}],[\"$\",\"td\",null,{\"children\":\"71.2\"}],[\"$\",\"td\",null,{\"children\":\"70.6\"}],[\"$\",\"td\",null,{\"children\":\"99.2%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"zen4-ultra\"}],[\"$\",\"td\",null,{\"children\":\"AIME 2024\"}],[\"$\",\"td\",null,{\"children\":\"94.7\"}],[\"$\",\"td\",null,{\"children\":\"93.6\"}],[\"$\",\"td\",null,{\"children\":\"98.8%\"}]]}]]}]]}]}]\n"])</script><script>self.__next_f.push([1,"2c:[\"$\",\"p\",null,{\"children\":[\"Average behavioral retention: \",[\"$\",\"strong\",null,{\"children\":\"99.3%\"}],\". The 0.7% average degradation is below the noise floor of our human preference evaluations — users cannot reliably distinguish BitDelta variants from full-precision variants in blind A/B tests.\"]}]\n2d:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"monosoup-svd-fallback-for-weak-checkpoints\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#monosoup-svd-fallback-for-weak-checkpoints\",\"className\":\"peer\",\"children\":\"MonoSoup: SVD Fallback for Weak Checkpoints\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n2e:[\"$\",\"p\",null,{\"children\":[[\"$\",\"strong\",null,{\"children\":\"Paper\"}],\" : arXiv:2602.09689\"]}]\n2f:[\"$\",\"p\",null,{\"children\":\"BitDelta works well when the delta is well-behaved (small, distributed, near-Laplace). Some fine-tuned checkpoints — particularly those from aggressive few-shot fine-tuning or noisy datasets — produce deltas that are large and spiky. In these cases, 1-bit quantization introduces perceptible degradation.\"}]\n30:[\"$\",\"p\",null,{\"children\":\"MonoSoup provides a complementary approach: instead of compressing the delta, decompose the full fine-tuned weight via SVD and keep only the top-k singular triplets:\"}]\n31:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" W_ft ≈ U_k Σ_k V_k^T\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n32:[\"$\",\"p\",null,{\"children\":\"where k is chosen to keep 95% of the Frobenius norm. This is not a delta compression technique — it operates on the single fine-tuned checkpoint directly. But for weak checkpoints where BitDelta degrades, MonoSoup recovers up to 8% of the lost behavioral accuracy at comparable memory cost.\"}]\n33:[\"$\",\"p\",null,{\"children\":\"In our pipeline: we try BitDelta first. If behavioral retention falls below 98.5% on our internal benchmark suite, we fall back to MonoSoup with k calibrated to budget.\"}]\n34:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"k-merge-edge-adapter-management\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#k-merge-edge-adapter-management\",\"className\":\"peer\",\"children\":\"K-Merge: Edge Adapter Management\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 "])</script><script>self.__next_f.push([1,"0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n35:[\"$\",\"p\",null,{\"children\":[[\"$\",\"strong\",null,{\"children\":\"Paper\"}],\" : arXiv:2510.13537\"]}]\n36:[\"$\",\"p\",null,{\"children\":\"The cloud serving stack above does not address edge deployment. A local user running zen-nano on a 16GB laptop cannot afford a delta cache for 14 variants — even at 1-bit compression, storing all nano variants would consume significant RAM.\"}]\n37:[\"$\",\"p\",null,{\"children\":[\"K-Merge addresses this with an \",[\"$\",\"strong\",null,{\"children\":\"online LoRA adapter pool under fixed storage budget\"}],\". The algorithm maintains a priority queue of adapters scored by utility:\"]}]\n38:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" utility(adapter_i) = request_frequency(i) × behavioral_gain(i) / storage_cost(i)\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]]}]}]}]\n39:[\"$\",\"p\",null,{\"children\":\"When the budget is exceeded, the lowest-utility adapter is evicted. Utility scores are updated online using exponential decay, so recently used adapters are preferred over historical ones.\"}]\n3a:[\"$\",\"p\",null,{\"children\":\"For a 16GB laptop with 4GB allocated to the adapter pool, K-Merge keeps 6-8 zen-nano variants hot simultaneously, with eviction latency of ~200ms to load a new adapter from local disk.\"}]\n3b:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"full-zen-serving-stack\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#full-zen-serving-stack\",\"className\":\"peer\",\"children\":\"Full Zen Serving Stack\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n"])</script><script>self.__next_f.push([1,"3c:[\"$\",\"$L44\",null,{\"className\":\"shiki shiki-themes github-light github-dark\",\"style\":{\"--shiki-light\":\"#24292e\",\"--shiki-dark\":\"#e1e4e8\",\"--shiki-light-bg\":\"#fff\",\"--shiki-dark-bg\":\"#24292e\"},\"tabIndex\":\"0\",\"icon\":\"\u003csvg viewBox=\\\"0 0 24 24\\\"\u003e\u003cpath d=\\\"M 6,1 C 4.354992,1 3,2.354992 3,4 v 16 c 0,1.645008 1.354992,3 3,3 h 12 c 1.645008,0 3,-1.354992 3,-3 V 8 7 A 1.0001,1.0001 0 0 0 20.707031,6.2929687 l -5,-5 A 1.0001,1.0001 0 0 0 15,1 h -1 z m 0,2 h 7 v 3 c 0,1.645008 1.354992,3 3,3 h 3 v 11 c 0,0.564129 -0.435871,1 -1,1 H 6 C 5.4358712,21 5,20.564129 5,20 V 4 C 5,3.4358712 5.4358712,3 6,3 Z M 15,3.4140625 18.585937,7 H 16 C 15.435871,7 15,6.5641288 15,6 Z\\\" fill=\\\"currentColor\\\" /\u003e\u003c/svg\u003e\",\"children\":[\"$\",\"$L45\",null,{\"children\":[\"$\",\"code\",null,{\"children\":[[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ┌──────────────────────────────┐\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ Request Router │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ (model ID → variant key) │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" └────────────┬─────────────────┘\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ┌────────────▼─────────────────┐\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ Delta Cache (Redis) │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ sign_bits + alpha per layer │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ ~9 GB per 72B variant │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" └────────────┬─────────────────┘\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ cache hit\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ┌────────────▼─────────────────┐\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ Shared Base Model (BF16) │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ zen-max 72B: 144 GB on A100s │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ zen-nano 0.6B: 1.2 GB │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" └────────────┬─────────────────┘\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ┌────────────▼─────────────────┐\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ Fused Reconstruction Kernel │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ W_ft = W_base + α·sign(Δ) │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ \u003c 1ms overhead per layer │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" └────────────┬─────────────────┘\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" ┌────────────▼─────────────────┐\"}]}],\"\\n\",[\"$\",\"span\",null,{\"className\":\"line\",\"children\":\"$L5d\"}],\"\\n\",\"$L5e\",\"\\n\",\"$L5f\",\"\\n\",\"$L60\"]}]}]}]\n"])</script><script>self.__next_f.push([1,"3d:[\"$\",\"p\",null,{\"children\":\"The architecture keeps one base model loaded per GPU cluster. All variants share it. The delta cache fits in Redis (NVMe-backed), loading on demand in under 50ms. In practice, our top-5 variants stay hot in Redis memory; the remaining variants load from NVMe on first request.\"}]\n3e:[\"$\",\"h2\",null,{\"className\":\"flex scroll-m-28 flex-row items-center gap-2\",\"id\":\"gpu-memory-reduction\",\"children\":[[\"$\",\"a\",null,{\"data-card\":\"\",\"href\":\"#gpu-memory-reduction\",\"className\":\"peer\",\"children\":\"GPU Memory Reduction\"}],[\"$\",\"svg\",null,{\"ref\":\"$undefined\",\"xmlns\":\"http://www.w3.org/2000/svg\",\"width\":24,\"height\":24,\"viewBox\":\"0 0 24 24\",\"fill\":\"none\",\"stroke\":\"currentColor\",\"strokeWidth\":2,\"strokeLinecap\":\"round\",\"strokeLinejoin\":\"round\",\"className\":\"lucide lucide-link size-3.5 shrink-0 text-fd-muted-foreground opacity-0 transition-opacity peer-hover:opacity-100\",\"aria-hidden\":true,\"children\":[[\"$\",\"path\",\"1cjeqo\",{\"d\":\"M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71\"}],[\"$\",\"path\",\"19qd67\",{\"d\":\"M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71\"}],\"$undefined\"]}]]}]\n3f:[\"$\",\"div\",null,{\"className\":\"relative overflow-auto prose-no-margin my-6\",\"children\":[\"$\",\"table\",null,{\"children\":[[\"$\",\"thead\",null,{\"children\":[\"$\",\"tr\",null,{\"children\":[[\"$\",\"th\",null,{\"children\":\"Scenario\"}],[\"$\",\"th\",null,{\"children\":\"Without BitDelta\"}],[\"$\",\"th\",null,{\"children\":\"With BitDelta\"}],[\"$\",\"th\",null,{\"children\":\"Savings\"}]]}]}],[\"$\",\"tbody\",null,{\"children\":[[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"3× zen-nano variants\"}],[\"$\",\"td\",null,{\"children\":\"3.6 GB\"}],[\"$\",\"td\",null,{\"children\":\"1.4 GB\"}],[\"$\",\"td\",null,{\"children\":\"61%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"5× zen4-max variants\"}],[\"$\",\"td\",null,{\"children\":\"300 GB\"}],[\"$\",\"td\",null,{\"children\":\"192 GB\"}],[\"$\",\"td\",null,{\"children\":\"36%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"2× zen-max variants\"}],[\"$\",\"td\",null,{\"children\":\"288 GB\"}],[\"$\",\"td\",null,{\"children\":\"162 GB\"}],[\"$\",\"td\",null,{\"children\":\"44%\"}]]}],[\"$\",\"tr\",null,{\"children\":[[\"$\",\"td\",null,{\"children\":\"Full 14-model catalog\"}],[\"$\",\"td\",null,{\"children\":\"~2.8 TB\"}],[\"$\",\"td\",null,{\"children\":\"~2.2 TB\"}],[\"$\",\"td\",null,{\"children\":\"21%\"}]]}]]}]]}]}]\n40:[\"$\",\"p\",null,{\"children\":\"The savings are most dramatic at the smaller scales where we have many more behavioral variants. For the ultra-scale models (zen4-ultra 1T+), a single checkpoint dominates, and BitDelta’s contribution is smaller — but MonoSoup and K-Merge become more relevant for edge quantization.\"}]\n41:[\"$\",\"p\",null,{\"children\":\"The combination of BitDelta for cloud serving, MonoSoup for quality recovery, and K-Merge for edge devices gives us a coherent three-tier compression story across the full Zen catalog.\"}]\n42:[\"$\",\"hr\",null,{}]\n43:[\"$\",\"p\",null,{\"children\":[\"$\",\"em\",null,{\"children\":\"Zen LM is a joint initiative of Hanzo AI Inc. (Techstars ‘17) and Zoo Labs Foundation (501c3).\"}]}]\n"])</script><script>self.__next_f.push([1,"47:[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" (delta \"}]\n48:[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"\u003e\"}]\n49:[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 0\"}]\n4a:[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\").to(torch.uint8) \"}]\n4b:[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6A737D\",\"--shiki-dark\":\"#6A737D\"},\"children\":\"# 1 = positive, 0 = negative\"}]\n4c:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" return\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" sign_bits, alpha\"}]]}]\n4d:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" \"}]}]\n4e:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" def\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6F42C1\",\"--shiki-dark\":\"#B392F0\"},\"children\":\" reconstruct_weight\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\"(W_base: torch.Tensor, sign_bits: torch.Tensor, alpha: torch.Tensor) -\u003e torch.Tensor:\"}]]}]\n4f:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" \\\"\\\"\\\"Reconstruct fine-tuned weight from base + compressed delta.\\\"\\\"\\\"\"}]}]\n50:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" signs \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" sign_bits.float() \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"*\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 2\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" -\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 1\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6A737D\",\"--shiki-dark\":\"#6A737D\"},\"children\":\" # map {0,1} → {-1,+1}\"}]]}]\n51:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" return\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" W_base \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"+\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" alpha \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"*\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" signs\"}]]}]\n52:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" \"}]}]\n53:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" def\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6F42C1\",\"--shiki-dark\":\"#B392F0\"},\"children\":\" memory_savings\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\"(d_out: \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\"int\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\", d_in: \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\"int\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"c"])</script><script>self.__next_f.push([1,"hildren\":\") -\u003e \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\"dict\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\":\"}]]}]\n54:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" \\\"\\\"\\\"Compare memory usage: full delta vs BitDelta.\\\"\\\"\\\"\"}]}]\n55:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" full_bytes \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" d_out \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"*\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" d_in \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"*\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 2\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6A737D\",\"--shiki-dark\":\"#6A737D\"},\"children\":\" # bfloat16\"}]]}]\n56:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" bitdelta_bytes \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"=\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" d_out \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"*\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" d_in \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"//\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 8\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" +\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 4\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#6A737D\",\"--shiki-dark\":\"#6A737D\"},\"children\":\" # 1-bit + float32 scale\"}]]}]\n57:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\" return\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" {\"}]]}]\n58:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" 'full_delta_mb'\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\": full_bytes \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"/\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 1e6\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\",\"}]]}]\n59:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" 'bitdelta_mb'\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\": bitdelta_bytes \"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"/\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#005CC5\",\"--shiki-dark\":\"#79B8FF\"},\"children\":\" 1e6\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\",\"}]]}]\n5a:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#032F62\",\"--shiki-dark\":\"#9ECBFF\"},\"children\":\" 'compression_ratio'\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\": full_bytes \"}],[\"$\",\"span\",null,{\"style"])</script><script>self.__next_f.push([1,"\":{\"--shiki-light\":\"#D73A49\",\"--shiki-dark\":\"#F97583\"},\"children\":\"/\"}],[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" bitdelta_bytes,\"}]]}]\n5b:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" }\"}]}]\n5c:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"style\":{\"--shiki-light\":\"#24292E\",\"--shiki-dark\":\"#E1E4E8\"},\"children\":\" \"}]}]\n5d:[\"$\",\"span\",null,{\"children\":\" │ Inference Engine │\"}]\n5e:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" │ (vLLM with continuous batch) │\"}]}]\n5f:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" └──────────────────────────────┘\"}]}]\n60:[\"$\",\"span\",null,{\"className\":\"line\",\"children\":[\"$\",\"span\",null,{\"children\":\" \"}]}]\n"])</script><script>self.__next_f.push([1,"b:[[\"$\",\"meta\",\"0\",{\"charSet\":\"utf-8\"}],[\"$\",\"meta\",\"1\",{\"name\":\"viewport\",\"content\":\"width=device-width, initial-scale=1\"}],[\"$\",\"meta\",\"2\",{\"name\":\"theme-color\",\"media\":\"(prefers-color-scheme: dark)\",\"content\":\"#0A0A0A\"}],[\"$\",\"meta\",\"3\",{\"name\":\"theme-color\",\"media\":\"(prefers-color-scheme: light)\",\"content\":\"#fff\"}]]\n"])</script><script>self.__next_f.push([1,"9:null\nd:[[\"$\",\"title\",\"0\",{\"children\":\"BitDelta: 1-Bit Behavioral Compression Across the Zen Model Family — Zen LM Blog | Zen LM\"}],[\"$\",\"meta\",\"1\",{\"name\":\"description\",\"content\":\"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.\"}],[\"$\",\"meta\",\"2\",{\"property\":\"og:title\",\"content\":\"Zen LM - Open Foundation Models\"}],[\"$\",\"meta\",\"3\",{\"property\":\"og:description\",\"content\":\"Zen AI model family by Hanzo AI. 14 frontier models from 4B to 1T+ parameters for code, reasoning, vision, and multimodal tasks.\"}],[\"$\",\"meta\",\"4\",{\"property\":\"og:url\",\"content\":\"https://zenlm.org\"}],[\"$\",\"meta\",\"5\",{\"property\":\"og:site_name\",\"content\":\"Zen LM\"}],[\"$\",\"meta\",\"6\",{\"property\":\"og:type\",\"content\":\"website\"}],[\"$\",\"meta\",\"7\",{\"name\":\"twitter:card\",\"content\":\"summary_large_image\"}],[\"$\",\"meta\",\"8\",{\"name\":\"twitter:site\",\"content\":\"@zenlmorg\"}],[\"$\",\"meta\",\"9\",{\"name\":\"twitter:creator\",\"content\":\"@zenlmorg\"}],[\"$\",\"meta\",\"10\",{\"name\":\"twitter:title\",\"content\":\"Zen LM - Open Foundation Models\"}],[\"$\",\"meta\",\"11\",{\"name\":\"twitter:description\",\"content\":\"Zen AI model family by Hanzo AI. 14 frontier models from 4B to 1T+ parameters for code, reasoning, vision, and multimodal tasks.\"}]]\n"])</script><script>
(function(){
// After React renders the lotus emoji from RSC payload, walk the DOM
// and replace it with the enso SVG. Polls so subsequent React re-renders
// also get cleaned up.
function swap(){
// 1. Replace lotus + ' Zen LM' text inside <a> brand links
document.querySelectorAll('a').forEach(function(a){
if (a.dataset.zenLotusSwapped) return;
var t = (a.textContent || '');
if (!/.*Zen LM|Zen LM.*/.test(t)) return;
// Find the lotus text node specifically
var walker = document.createTreeWalker(a, NodeFilter.SHOW_TEXT, null);
var node;
while ((node = walker.nextNode())) {
if (node.nodeValue && node.nodeValue.indexOf('') !== -1) {
node.nodeValue = node.nodeValue.replace(/\s?/g, '');
}
}
// Inject enso SVG at start
if (!a.querySelector('svg[aria-label="Zen LM"]')) {
var ns = 'http://www.w3.org/2000/svg';
var svg = document.createElementNS(ns, 'svg');
svg.setAttribute('viewBox', '0 0 512 512');
svg.setAttribute('fill', 'none');
svg.setAttribute('width', '18');
svg.setAttribute('height', '18');
svg.setAttribute('aria-label', 'Zen LM');
svg.style.cssText = 'display:inline-block;vertical-align:-2px;margin-right:6px;flex-shrink:0';
var p = document.createElementNS(ns, 'path');
p.setAttribute('d', 'M 256 84 A 172 172 0 1 0 256 428 A 172 172 0 1 0 256 84 Z');
p.setAttribute('stroke', 'currentColor');
p.setAttribute('stroke-width', '26');
p.setAttribute('stroke-linecap', 'round');
p.setAttribute('stroke-linejoin', 'round');
svg.appendChild(p);
a.insertBefore(svg, a.firstChild);
}
a.dataset.zenLotusSwapped = '1';
});
}
function start(){ swap(); setInterval(swap, 600); }
if (document.readyState !== 'loading') start();
else document.addEventListener('DOMContentLoaded', start);
})();
</script>
<style id="zenlm-brand-anim">
@keyframes zenlm-brand-draw { to { stroke-dashoffset: 0; } }
svg.zenlm-brand-svg .zenlm-brand-path {
stroke-dasharray: 1100;
stroke-dashoffset: 1100;
animation: zenlm-brand-draw 1.4s cubic-bezier(.45,0,.2,1) 0.1s 1 forwards;
}
.zen-link-icon { display: inline-block; vertical-align: middle; flex-shrink: 0; }
</style></body></html>