Ornith‑1.0 is a new family of open‑source, self‑improving models for agentic coding tasks, ranging from a 9B dense model for edge deployment to a 397B mixture‑of‑experts (MoE) flagship. Built on Gemma 4 and Qwen 3.5, Ornith‑1.0 jointly learns both task solutions and the scaffolding (harnesses) that guide those solutions, enabling it to discover more effective search trajectories.

Key Highlights

• Performance

  • 397B version matches/exceeds Claude Opus 4.7 on Terminal‑Bench 2.1 (77.5 vs 70.3) and SWE‑Bench Verified (82.4 vs 80.8).
  • 35B variant outperforms similarly sized models and even beats larger Qwen 3.5‑397B on Terminal‑Bench.
  • 9B edge model delivers strong results (43.1 on Terminal‑Bench, 69.4 on SWE‑Bench) comparable to much larger models.

• Self‑Improving Training

  • Each RL step first refines the scaffold, then generates a solution conditioned on it; rewards propagate to both stages, allowing scaffolds to evolve automatically.

• Safety & Reward‑Hacking Mitigations

  1. Fixed outer trust boundary (environment, tool surface, test isolation).
  2. Deterministic monitor flags forbidden actions, assigning zero reward.
  3. Frozen LLM judge acts as a veto on top of the verifier.

• Asynchronous RL

  • Uses pipeline‑RL with staleness weighting to handle long rollouts, down‑weighting older tokens.

https://deep-reinforce.com/ornith_1_0.html