release
Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding
Ornith-1.0 provides a powerful, locally runnable option for developers building agentic coding workflows, reducing reliance on proprietary APIs and enabling complex multi-step automation.
What happened
DeepReinforce has released Ornith-1.0, a family of open-weights coding models under the MIT license, according to Simon Willison. Built by fine-tuning Gemma 4 and Qwen 3.5, the models range from 9B to 397B parameters and achieve state-of-the-art results on coding benchmarks among open-source models of similar size. Willison tested the 35B MoE variant via LM Studio and Pi, noting its ability to handle multi-step tool-calling tasks, such as locating code in a codebase and generating images. Despite some mangled output in a pelican drawing, the model demonstrated proficiency in agentic coding workflows. DeepReinforce appears to be a new entrant, with their earliest paper from June 2025 on CUDA optimization. The licensing of the base models (Apache 2.0) allows for this combination, addressing past concerns with Gemma's terms.
Key takeaways
- Ornith-1.0 is an MIT-licensed open model family from DeepReinforce, based on Gemma 4 and Qwen 3.5.
- Sizes include 9B, 31B, 35B MoE, and 397B MoE, with state-of-the-art coding benchmark results.
- Simon Willison tested the 35B variant via LM Studio and Pi, finding it capable of complex agentic tasks like code navigation and tool use.
- DeepReinforce's earlier research includes CUDA-L1 (June 2025) on contrastive reinforcement learning for CUDA optimization.
- The model is compatible with existing tools like LM Studio for local deployment.
Why it matters
Ornith-1.0 provides a powerful, locally runnable option for developers building agentic coding workflows, reducing reliance on proprietary APIs and enabling complex multi-step automation.
This is an original editorial digest by AI Workflow Pro. Full reporting at the source:
Read the original on Simon WillisonMore AI news
All news →
Join the AI Workflow Pro Community