Zen Models
Language · Code · Vision · Audio
45+ open-weight models from 0.6B to 1T+, built by Zoo Labs Foundation and served on the Hanzo API. Zen MoDE (Mixture of Diverse Experts) architecture. All models on HuggingFace.
Full Model Catalog
45+ models · Available via API and HuggingFace
What you get by using ours
Same request shape
Change the model name in a call you already wrote. Nothing else about the request changes.
Weights you can take
Most Zen models are published on HuggingFace, so you can serve them yourself. A few are API-only; the catalog above marks which.
0.6B to frontier
Small enough to run on a laptop, or large enough to reason. Text, code, vision, audio, image and video, one API.
Sparse where it counts
The Mixture-of-Diverse-Experts models hold a large parameter count but activate a small slice of it per token, so capacity is not what you pay for.
Pick by what you are doing
Different models, one endpoint. The name in the request is the only thing that changes
Language
Chat, reasoning, retrieval
Code
A million tokens of it
Vision
Read a screen, draw a thing
Audio
Listen, speak, compose
Zen 5 Ultra
A 2T-parameter MoDE, training now inside NVIDIA TEE confidential compute on hanzo.network. The point of doing it that way is that the run attests to itself: what data went in and what came out is checkable afterwards rather than taken on our word.
Why they are named this way
These models are built on eight engineering principles drawn from the 64 hexagrams of the I-Ching — orthogonality, smallness, completeness, clarity and composability among them. Ancient pattern language for systems that last.
易經 · Explore the Full PhilosophyCall one and see
Start with zen5-mini because it is fast and cheap, and move up only where a harder question needs it.