Ornith-1.0 is an open coding model that scaffolds itself

Simon Willison has been testing Ornith-1.0, a new family of open-weights coding models from a group called DeepReinforce, and his verdict is that the open-source side now has another serious agentic coder. The release is unusually broad: four variants, from a 9B dense model up to a 397B mixture-of-experts, with 31B dense and 35B MoE in between, all under an MIT license. The models are built on top of Gemma 4 and Qwen 3.5, both Apache 2.0, so the licensing stays clean for people who want to build on the result. DeepReinforce claims state-of-the-art results among open models of similar size on coding benchmarks.

The interesting word is self-scaffolding. Rather than relying on an external harness to break a task into steps, Ornith structures its own reasoning across a long sequence of tool calls and manages that flow itself. Willison ran the 35B variant locally from a 20GB GGUF quantization and saw about 103 tokens per second, and he reports it handled real work like navigating an unfamiliar code repository and chaining many tool calls without losing the thread. The full write-up is here.

Why it matters

If you run coding agents and want to keep them local or avoid per-token API costs, an MIT-licensed model that scaffolds its own tool use is worth downloading and pointing at your own repo before you trust the benchmark claims.