DeepReinforce released Ornith-1.0 family on Hugging Face June 25, 2026. 397B MoE flagship for agentic coding with reasoning blocks and tool-call parsing. Benchmarks: Terminal-Bench 2.1, SWE-Bench, NL2Repo, OpenClaw.
Self-improving training: RL learns solution rollouts and scaffolds jointly. OpenAI-compatible serving recipes documented. Collection includes 8 model artifacts across four sizes.
MIT license — commercial use permitted without legal friction of restricted open-weight licenses.