DeepReinforce has released Ornith-1.0, the first open-weight model explicitly designed for coding agency. Available in four sizes (9B to 397B parameters, dense and MoE), it’s built on top of Gemma 4 and Qwen 3.5, with an MIT license. The model learns to autonomously build the scaffold — the search structure and intermediate steps — that guides solving complex coding tasks. It does this through reinforcement learning: it doesn’t just generate code, but the path to get there.

On agentic benchmarks (SWE-bench Verified, Terminal-Bench 2.1, NL2Repo), Ornith-1.0 outperforms comparably sized open models. The 35B version, for instance, solves 75.6% of SWE-bench Verified tasks against 70% for Qwen3.5-35B (June 2026). The 397B holds its ground against frontier proprietary models.

Why it matters to you: until now, high-performing coding agents required either proprietary APIs or open models adapted in-house. Ornith-1.0 is the first release designed to run locally, with a license that has no regional or commercial restrictions, and scales from laptops (the 9B runs quantized in GGUF) to clusters. If you’re building tools that automate coding on real codebases — refactoring, repo analysis, test generation — you now have an open alternative purpose-built for that use case.

Where to look: the GitHub repository has benchmarks and weight links. A public test session with the 35B quantized in LM Studio shows the model navigating a Datasette repo with multiple tool calls and drawing a pelican in SVG (at 103 tokens/s). DeepReinforce is a new player: their first paper came out in June 2025.