Ornith-1.0: self-improving open-source models for agentic coding
Scored daily by a customisable AI persona to surface the most relevant engineering leadership news.
Self-improving open-source agentic coding model, directly matches AI agent and open-source interests.
Ornith-1.0 is an MIT-licensed, self-improving open-source model family for agentic coding, post-trained on Gemma 4 and Qwen 3.5 across 9B, 31B, 35B-MoE, and 397B-MoE sizes. It uses reinforcement learning to jointly optimize both solution rollouts and the scaffold driving them, achieving state-of-the-art results on Terminal-Bench 2.1 (77.5 on Terminus-2), SWE-bench Verified (82.4), and NL2Repo (48.2) for the 397B variant, often matching or exceeding proprietary models like Claude Opus 4.7. The models are reasoning models with tool-call support, requiring recent vLLM for serving.
deepreinforce-ai