mirror of
https://github.com/hanzo-docs/docs.git
synced 2026-10-28 07:09:50 +00:00
zeekay
4bd5378921
- packages/zen-models: rename package to @zenlm/models, update all source comments - zen-docs: fix "Mixture of Distilled Experts" → "Mixture of Diverse Experts" across all blog posts and docs content (models.mdx, index.mdx, zen-mode-architecture.mdx, zen4-ultra.mdx, zen5-release.mdx, drop-upcycling-zen-mode.mdx) - apps/docs: fix brand copy on home page
66 lines
2.7 KiB
Plaintext
66 lines
2.7 KiB
Plaintext
---
|
|
title: Introducing Zen 5 — Our Most Capable Model Yet
|
|
description: Zen 5 brings 1T+ parameter scale, 2M context windows, and state-of-the-art performance across reasoning, code, and multimodal tasks.
|
|
author: Zen LM Team
|
|
date: "2026-03-01"
|
|
tags: [release, zen5, flagship]
|
|
---
|
|
|
|
We're thrilled to announce the **Zen 5** family — our most capable generation of foundation models, pushing the frontier of what open models can achieve.
|
|
|
|
## What's New in Zen 5
|
|
|
|
Zen 5 represents a major leap forward across every dimension that matters:
|
|
|
|
### Scale and Architecture
|
|
|
|
Zen 5 is built on the **Zen MoDE** (Mixture of Diverse Experts) architecture — a framework that combines the best of sparse expert routing with distillation from multiple frontier models. At 1T+ total parameters with intelligent expert activation, Zen 5 achieves flagship intelligence without the inference cost of a dense trillion-parameter model.
|
|
|
|
### 2M Context Window
|
|
|
|
Zen 5 supports **2 million tokens** of context — enough to process entire codebases, multi-hour transcripts, or hundreds of research papers in a single request. This unlocks agentic workflows that simply weren't possible before.
|
|
|
|
### The Zen 5 Family
|
|
|
|
| Model | Parameters | Context | Strengths |
|
|
|-------|-----------|---------|-----------|
|
|
| zen5 | 1T+ MoE | 2M | Flagship reasoning, all tasks |
|
|
| zen5-pro | 480B MoE | 512K | High capability, cost-efficient |
|
|
| zen5-ultra | 1T+ MoE | 2M | Maximum reasoning depth |
|
|
| zen5-mini | 14B | 128K | Fast, lightweight inference |
|
|
| zen5-max | 1T+ MoE | 2M | Enterprise scale |
|
|
| zen5-coder | 1T+ MoE | 2M | Code generation, software engineering |
|
|
|
|
## Benchmark Performance
|
|
|
|
Zen 5 achieves state-of-the-art results on major benchmarks:
|
|
|
|
- **MMLU**: 92.1% (5-shot)
|
|
- **HumanEval**: 96.3% (pass@1)
|
|
- **MATH**: 91.7%
|
|
- **GPQA Diamond**: 75.2%
|
|
|
|
## Open Weights
|
|
|
|
True to our commitment to open AI, Zen 5 weights are available on [HuggingFace](https://huggingface.co/zenlm/zen5) under the Zen Open License. The model runs efficiently in 4-bit quantization on consumer hardware using our optimized inference stack.
|
|
|
|
## Getting Started
|
|
|
|
```bash
|
|
# Install the Zen models catalog
|
|
npm install @zenlm/models
|
|
|
|
# Use via the Hanzo API
|
|
curl https://api.hanzo.ai/v1/chat/completions \
|
|
-H "Authorization: Bearer $HANZO_API_KEY" \
|
|
-d '{"model": "zen5", "messages": [{"role": "user", "content": "Hello!"}]}'
|
|
```
|
|
|
|
Or try it instantly at [hanzo.chat](https://hanzo.chat?model=zen5).
|
|
|
|
## What's Next
|
|
|
|
We're actively working on Zen 5 fine-tunes for specialized domains — medicine, law, materials science — and expanding multimodal capabilities. The next release will bring native video understanding and real-time audio processing to the Zen 5 family.
|
|
|
|
Stay tuned.
|