Zen Max: 671B Reasoning Model

Zen Max is a 671B mixture-of-experts reasoning model. It is currently the most capable model in the Zen MoDE lineup, and one of the highest-performing open-weight models available on any benchmark.

It is also abliterated.

Architecture

671B total parameters, 384 experts, 8 active per forward pass. At inference time, Zen Max activates approximately 14B parameters per token -- the compute profile of a mid-size dense model with the knowledge and capability of a 671B one.

This architecture makes Zen Max deployable at frontier quality on configurations that would be impractical for a dense 671B model. 8x H100 SXM handles the full-precision model. 4x H100 with FP8 quantization works for production throughput.

Benchmark Results

The AIME 2025 result of 99.1% is the number to look at. AIME is the American Invitational Mathematics Examination -- problems that a small fraction of mathematically gifted high school students solve correctly. Getting 99.1% requires genuine mathematical reasoning, not pattern matching.

SWE-Bench Verified at 71.3% means Zen Max resolves 71.3% of real GitHub issues from open-source projects when given the repository and the issue description. This is an agentic task requiring understanding of existing code, planning a fix, and implementing it correctly.

Extended Thinking

256K context window with extended reasoning enabled. For complex multi-step problems, Zen Max can think for as long as the problem requires before generating output. The model allocates compute to thinking proportional to problem difficulty -- short answers to simple questions, extended reasoning chains for research and hard math.

Agentic Tool Use

Zen Max handles 200-300 sequential tool calls in a single agentic task. This is not a typical generation loop -- it is deep autonomous operation: the model issues tool calls, observes results, updates its plan, and continues. Tasks that require browsing dozens of web pages, running hundreds of code experiments, or traversing a large codebase end-to-end are within scope.

Abliteration

Zen Max ships with an abliterated base. Abliteration is directional ablation of refusal behavior from the model's residual stream. The refusal directions -- the learned internal representations that cause a model to decline requests -- are identified and removed from the weights before release.

The result is a model that reasons about any topic without built-in filtering. For researchers, security professionals, and applications where full capability access matters, this is the model.

Abliteration does not degrade benchmark performance. AIME, SWE-Bench, and GPQA scores are measured on the abliterated model.

Get Zen Max

• HuggingFace: huggingface.co/zenlm
• Hanzo Cloud API: api.hanzo.ai/v1/chat/completions -- model zen-max
• Zen LM: zenlm.org -- deployment guides, hardware requirements

64 SafeTensor shards. Download with hf download zenlm/zen-max.

---

Zach Kelling is the founder of Hanzo AI, Techstars '17.