Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

Main idea

Ornith-1.0 reaches state-of-the-art performance among similarly sized open-source models on coding benchmarks and handles multi-step tool-use sequences in practice. Willison's hands-on tests confirm the model can reliably navigate a real repository (Datasette).

Context

DeepReinforce is a relatively unknown player (the only public trace is a June 2025 paper on CUDA optimization), so an MIT-licensed frontier-grade coding model release is a surprise. It arrives as the community is searching for local alternatives to Claude Code and Cursor. Willison tested it on release day via LM Studio.

Why it matters

For practitioners this is another strong open-weights option for local coding agents without subscription lock-in. The MIT license removes commercial barriers and LM Studio compatibility lowers the technical bar for teams without their own inference infrastructure.

Details / arguments

• Variants: 9B Dense, 31B Dense, 35B MoE, 397B MoE
• Built on Gemma 4 and Qwen 3.5 (both Apache 2.0)
• MIT license (commercially usable)
• Willison measured 103 tokens/sec generating ASCII art
• 35B GGUF variant is 20 GB and runs locally in LM Studio