DeepReinforce released Ornith-1.0 June 25, 2026 — four MIT-licensed sizes (9B, 31B, 35B-MoE, 397B-MoE) on Hugging Face, post-trained on Gemma 4 and Qwen 3.5.
Most coding agents use fixed human-designed harnesses. Ornith-1.0 learns its own scaffold during RL, jointly optimizing harness and solution. 397B tops Claude Opus 4.7 on headline benchmarks but trails Opus 4.8 and GLM-5.2-744B.
Anti-reward-hacking: fixed trust boundary (immutable outer layer), deterministic monitor (zero reward for reading withheld tests), frozen LLM judge as veto on top of verifier.