DeepReinforce released Ornith-1.0 on June 25, 2026 — a self-improving family of open-source models for agentic coding. Four sizes: 9B Dense, 31B Dense, 35B MoE, 397B MoE. Built on Gemma 4 and Qwen 3.5 foundations under MIT license on Hugging Face.
Key innovation: self-scaffolding RL framework. Instead of fixed human-designed harnesses, Ornith-1.0 learns to generate both solution rollouts and task-specific harnesses. Each RL step: (1) propose refined scaffold given task and prior scaffold; (2) generate solution conditioned on scaffold. Reward propagates to both stages.
397B flagship claims 82.4 SWE-Bench Verified and 77.5 Terminal-Bench 2.1. Uses asynchronous pipeline-RL with token-level GRPO loss. Three-layer anti-reward-hacking: fixed trust boundary, deterministic monitor, frozen LLM judge veto.