Ornith-1.0-35B brings a mid-size MoE to agentic coding

A new entry in the Ornith-1.0 family has landed on Hugging Face: Ornith-1.0-35B, a mixture-of-experts model aimed squarely at agentic coding workloads. Published under the permissive MIT license, it slots in as the mid-size option in a lineup built around what its makers describe as a self-scaffolding approach to code generation.

The headline number is 35 billion parameters in a mixture-of-experts configuration, which means only a fraction of those weights activate on any given token. That design lets a model reach for the reasoning capacity of a larger network while keeping inference costs closer to a smaller dense one — a practical tradeoff for developers who want capable coding assistance without the hardware bill of a frontier system.

Why it matters

Agentic coding — where a model plans, writes, and iterates on code across multiple steps rather than answering a single prompt — has become one of the most competitive corners of open-weights AI. A permissively licensed MoE at this scale is notable for a few reasons:

• The MIT license places minimal restrictions on commercial and derivative use.
• The MoE architecture makes self-hosting more approachable than a comparably capable dense model.
• A mid-size variant gives teams a middle ground between lightweight and heavyweight options in the same family.

As an initial 1.0 release, Ornith-1.0-35B arrives without the extensive benchmark history of established coding models, and the model card is the best source for configuration details and intended use. For teams evaluating open coding assistants, it's worth watching how the self-scaffolding claims hold up against real-world repositories.