Introducing Zen LM: Open Frontier Models from Hanzo AI and Zoo Labs — Zen LM Blog

GITHUB HUGGING FACE TRY ZEN CHAT

Today we are announcing Zen LM — a family of open frontier models co-developed by Hanzo AI and Zoo Labs Foundation. This release marks the public launch of the Zen model catalog: 94+ models spanning text, vision, audio, and code, all built on our Zen MoDE (Mixture of Distilled Experts) architecture.

Why We Built It

Modern AI infrastructure concentrates capability in a small number of proprietary systems. The models that power the most demanding production workloads are closed, expensive, and opaque. We believe frontier capability should be accessible — not as a service you pay per token, but as open weights you can run, inspect, fine-tune, and deploy on your own infrastructure.

Hanzo AI has spent years building AI infrastructure used by thousands of developers. Zoo Labs Foundation has led decentralized AI research through ZIPs (Zoo Improvement Proposals) and the PoAI (Proof of AI) protocol. Zen LM is the convergence of that work: production-grade models, open weights, governed by the community.

The Zen MoDE Architecture

Zen MoDE stands for Mixture of Distilled Experts. Every Zen model is built on this architecture:

Expert routing : Each token is processed by a small subset of specialized expert networks
Distillation from large teachers : Experts are initialized from, and continuously refined against, larger teacher models
Efficient inference : Only a fraction of total parameters activate per token, making large models economically viable

The result: Zen models deliver capability competitive with much larger dense models at a fraction of the inference cost.

For a deep technical treatment see our architecture post: Zen MoDE Architecture.

The Catalog

The Zen family covers every major use case:

Model	Parameters	Use Case
Zen4 Ultra	480B (35B active)	Maximum capability, frontier tasks
Zen Max	72B	General enterprise use
Zen4 Pro	32B (22B active)	Balanced capability/cost
Zen4 Flash	7B (3B active)	Low-latency production
Zen4 Coder	480B (35B active)	Code generation, agentic coding
Zen Omni	32B	Vision + text + audio
Zen VL	72B	Image understanding, OCR
Zen Nano	0.6B	On-device inference
Zen Embedding	7680-dim	Semantic search, RAG
Zen Guard	3B	Safety classification

All general-purpose models are released under Apache-2.0. Safety models use a more restrictive license to prevent adversarial use.

How We Train

Zen models are trained on the Zoo Compute Network — a decentralized GPU cluster funded through the Zoo Labs Foundation treasury and governed by ZIPs. Training details:

Data : 15T+ tokens, high-quality filtered corpus with documented provenance
Compute : Distributed across heterogeneous H100/A100 clusters
Alignment : GRPO (Group Relative Policy Optimization) for instruction following
Continuous improvement : ASO (Active Semantic Optimization, HIP-002) feeds production signals back into training

Open Weights Commitment

We release weights, not just APIs. Every Zen model is available on Hugging Face under hanzoai/. You can:

Download and run locally with transformers, vLLM, or llama.cpp
Fine-tune on your own data
Deploy on any infrastructure
Inspect weights, tokenizer, and training configuration

We believe open weights is the only credible commitment to openness. APIs can be taken down. Weights are permanent.

What’s Next

The Zen4 generation is live today. We are already training Zen5, with improvements to:

Long-context reasoning (targeting 2M token context)
Multimodal integration (tighter vision-text-audio coupling)
Agentic reliability (reduced hallucination in tool-use chains)
Distillation efficiency (more capability per active parameter)

Follow along at zenlm.org, on GitHub, and on Hugging Face.

Zen LM is a joint initiative of Hanzo AI Inc. (Techstars ‘17) and Zoo Labs Foundation (501c3).