mirror of
https://github.com/zenlm/zen-docs.git
synced 2026-10-30 14:58:13 +00:00
zeekay
8ef8f3e801
Updated 8 existing stubs and added 4 new model pages: - zen-nano: 0.6B edge model specs, throughput table, benchmarks - zen-scribe: 4B writing model, multilingual, use cases - zen-omni: 30B hypermodal Thinker-Talker MoE, speech-to-speech - zen-vl: 4B/8B/30B VL family with function calling and agent variants - zen-embedding: MTEB #1 multilingual (70.58), MRL support - zen-coder: 4B/31B/480B MoE family, FIM, 358 languages - zen-designer: 235B VL MoE, document parsing, GUI interaction - zen-max: 1T+ MoE, 256K context, 200+ tool calls, heavy mode - zen-pro: 8B instruct/thinking/agent variants - zen-reranker: 8B dual-task, native 7680-dim for DSO - zen-translator: real-time translation pipeline with voice cloning - zen-guard: removed upstream model reference No upstream model names in any page (Zen MoDE branding throughout).
112 lines
3.4 KiB
Plaintext
112 lines
3.4 KiB
Plaintext
---
|
|
title: Zen Max
|
|
description: 1T+ reasoning-first MoE model with 256K context, extended thinking, and 200+ sequential tool calls
|
|
---
|
|
|
|
# Zen Max
|
|
|
|
Zen Max is the largest model in the Zen family — a 1T+ reasoning-first language model built for test-time scaling through extended thinking and tool-calling. With 32B active parameters via MoE sparse activation and up to 128K thinking tokens per step, it executes 200-300 sequential tool calls without human intervention for agentic reasoning, search, coding, and mathematical problem solving.
|
|
|
|
## Specifications
|
|
|
|
| Property | Value |
|
|
|----------|-------|
|
|
| Total Parameters | 1.04T (1,044B) |
|
|
| Active Parameters | 32B (MoE, 384 experts, top-8 routing) |
|
|
| Architecture | Zen MoDE (MoE) |
|
|
| Context Window | 256K tokens |
|
|
| Thinking Capacity | 96K-128K tokens per step |
|
|
| License | Apache 2.0 |
|
|
| Formats | SafeTensors (BF16), INT4 (native QAT) |
|
|
|
|
## Key Capabilities
|
|
|
|
- **Agentic reasoning** — Extended chain-of-thought with `<think>` tags, up to 128K thinking tokens
|
|
- **Agentic search** — 200-300 sequential tool calls, real-world information gathering
|
|
- **Agentic coding** — Multi-language support, terminal automation, component-heavy web development
|
|
- **Mathematical reasoning** — PhD-level math with Python tool integration
|
|
- **Heavy Mode** — 8 parallel reasoning trajectories with reflective aggregation
|
|
|
|
## Quick Start
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("zenlm/zen-max")
|
|
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-max")
|
|
|
|
messages = [
|
|
{
|
|
"role": "user",
|
|
"content": "Research and analyze the latest developments in quantum computing, then write a comprehensive report."
|
|
}
|
|
]
|
|
|
|
# Model will think, search, browse, and synthesize autonomously
|
|
response = model.chat(tokenizer, messages, thinking_budget=128000, max_tool_calls=300)
|
|
```
|
|
|
|
## Use Cases
|
|
|
|
- Frontier-difficulty reasoning and research synthesis
|
|
- Large-scale software engineering (SWE-Bench: 71.3%)
|
|
- Mathematical olympiad problems (AIME 2025: 99.1%)
|
|
- Complex agentic workflows requiring hundreds of tool calls
|
|
- Expert-level domain Q&A (GPQA-Diamond: 84.5%)
|
|
|
|
## Benchmark Performance
|
|
|
|
### Reasoning
|
|
|
|
| Benchmark | Score |
|
|
|-----------|-------|
|
|
| HLE (with tools) | 44.9% |
|
|
| AIME 2025 (with Python) | 99.1% |
|
|
| HMMT 2025 (with Python) | 95.1% |
|
|
| IMO-AnswerBench | 78.6% |
|
|
| GPQA-Diamond | 84.5% |
|
|
|
|
### Agentic Search
|
|
|
|
| Benchmark | Score |
|
|
|-----------|-------|
|
|
| BrowseComp | 60.2% |
|
|
| Frames | 87.0% |
|
|
|
|
### Coding
|
|
|
|
| Benchmark | Score |
|
|
|-----------|-------|
|
|
| SWE-Bench Verified | 71.3% |
|
|
| LiveCodeBench v6 | 83.1% |
|
|
| Terminal-Bench | 47.1% |
|
|
|
|
## Formats
|
|
|
|
| Format | Size | Notes |
|
|
|--------|------|-------|
|
|
| SafeTensors (BF16) | ~370 GB | 62 shards, full precision |
|
|
| INT4 (native QAT) | ~185 GB | 2x speed, minimal quality loss |
|
|
|
|
## Hardware Requirements
|
|
|
|
| Setup | VRAM | Speed |
|
|
|-------|------|-------|
|
|
| Budget (1x 24GB + 256GB RAM) | 280 GB combined | ~1-2 tokens/s |
|
|
| High performance (4x A100 80GB) | 320 GB VRAM | 5+ tokens/s |
|
|
| Optimal (8x A100 80GB) | 640 GB VRAM | Maximum throughput |
|
|
|
|
## Thinking Budget
|
|
|
|
| Mode | Thinking Tokens | Use When |
|
|
|------|----------------|----------|
|
|
| Low | 32K | Fast responses |
|
|
| Medium | 96K | Balanced reasoning |
|
|
| High | 128K | Complex problems |
|
|
| Heavy | 8 x 128K | Maximum accuracy |
|
|
|
|
## Get Started
|
|
|
|
- [HuggingFace](https://huggingface.co/zenlm/zen-max)
|
|
- [GitHub](https://github.com/zenlm/zen-max)
|