Three Organizations, One Model Family

The Zen model family is the product of three organizations with distinct but complementary roles. Understanding how they fit together explains why Zen exists as open weights rather than a proprietary service.

Hanzo AI — Infrastructure and Scale

hanzo.ai | AI infrastructure | Techstars '17

Hanzo AI builds the infrastructure layer: the LLM gateway, the API, the tooling developers actually use in production. Since 2017, Hanzo has been operating at scale — routing model traffic, managing provider fallback, handling the operational complexity that comes with running AI systems in production environments.

Hanzo productizes Zen models for real-world use:

• console.hanzo.ai: API access with an OpenAI-compatible endpoint, per-token billing, enterprise SLA
• hanzo.chat: Consumer interface — try any Zen model free, no account required
• LLM Gateway (api.hanzo.ai): Unified proxy routing to Zen models and 100+ third-party providers

Hanzo's position in the infrastructure stack means every model improvement has a direct path to production. The feedback loop runs continuously: production signals inform training, training improves the models, better models improve the product.

Zoo Labs Foundation — Research and Governance

zoo.ngo | 501(c)(3) | Open AI research

Zoo Labs Foundation is a nonprofit research organization focused on open AI and decentralized science (DeSci). The Foundation does the research work that makes Zen possible:

• Training experiments: All model training happens under Foundation governance, with open logs, open data provenance documentation, and published negative results
• ZIPs governance: Major decisions — architecture choices, training methodology changes, data standards — go through the Zoo Improvement Proposals process, open for community input
• Open infrastructure: The Zoo Gym (training infrastructure), evaluation harnesses, and data curation tools are published under open licenses
• DeSci research: Decentralized science coordination, including federated training experiments and community-run evaluations

The Foundation structure matters. A nonprofit with published governance cannot quietly close its models, change its license, or stop publishing weights without violating its own charter. That permanence is the point.

Key protocols developed under Foundation governance:

• DSO (Decentralized Semantic Optimization, ZIP-001): Federated fine-tuning across distributed participants
• PoAI (Proof of AI, ZIP-002): Verifiable inference for cryptographic attestation of model outputs
• BitDelta (ZIP-007): Behavioral compression for efficient weight updates

Lux Partners Limited — Compute and Settlement

Decentralized training at the scale required to build frontier models requires more than cloud credits. Lux Partners provides the compute coordination and settlement infrastructure layer that makes distributed training economically viable.

What Lux enables:

• Distributed compute: Coordinating training across heterogeneous GPU clusters — H100s, A100s, and consumer hardware in the Zoo Compute Network
• Settlement infrastructure: Accounting for compute contributions, distributing rewards, and maintaining economic integrity across participants
• Blockchain-AI convergence: Lux's settlement layer is purpose-built for AI workloads, not adapted from general-purpose financial infrastructure

The combination of Lux's compute layer with Zoo Foundation's decentralized training protocols means Zen models can be trained across a genuinely distributed network rather than a single centralized data center.

How It Works in Practice

The division of responsibility follows naturally from each organization's core competency:

Zoo Foundation runs the research. Training experiments, architecture proposals, data curation, evaluation. All of it happens in public, under ZIPs governance, with the community able to participate, review, and challenge decisions.

Hanzo AI productizes and serves. Once weights are ready, Hanzo handles the infrastructure: quantization, serving optimization, API design, developer tooling, the consumer interface. Hanzo's operational experience means models ship into production with known performance characteristics, not just benchmark numbers.

Lux provides the compute layer. Distributed training coordination, compute settlement, and the decentralized infrastructure that lets anyone contribute resources to the network and be compensated fairly.

All Weights, Apache 2.0

The three-org structure produces one outcome worth emphasizing: every general-purpose Zen model ships under Apache 2.0.

This is not a license designed to sound open while retaining control. Apache 2.0 means:

• Commercial use permitted, no royalty
• Derivative works permitted
• No user restriction clauses
• No usage reporting requirements

The weights are available on huggingface.co/zenlm. Download them. Run them. Build on them. You do not need to ask permission.

Why This Structure

Building frontier models requires capital, compute, and sustained research effort. Building open frontier models requires all of that plus a structural commitment to openness that survives commercial pressure.

A single for-profit company building open models faces a straightforward conflict: openness costs, proprietary systems profit. Many "open" model releases have quietly added restrictions, delayed weight releases, or stopped releasing weights altogether when commercial interests aligned against it.

The three-org structure distributes that tension:

• The Foundation's nonprofit charter protects the open research mandate
• Hanzo's commercial position funds infrastructure and keeps the models in production use
• Lux's decentralized compute layer removes single-provider dependence

No single organization controls Zen. The weights, once released, are permanent. The governance is public. The training code is open.

Get Involved

• Use the models: hanzo.chat (free) or console.hanzo.ai (API)
• Download weights: huggingface.co/zenlm
• Contribute to research: zoo.ngo — submit ZIPs, join working groups, contribute compute
• Build on the infrastructure: github.com/hanzoai — training gym, evaluation tools, model code

---

Zen LM is a joint initiative of Hanzo AI Inc. (Techstars '17), Zoo Labs Foundation (501c3), and Lux Partners Limited.