9B dense, 35B MoE, 397B MoE publicly confirmed; official release page also describes a 31B dense variant
262,144 tokens
MIT
Official open-source Ornith-1.0 family release focused on agentic coding, built on top of pretrained Gemma 4 and Qwen 3.5 backbones and positioned around a self-improving reinforcement-learning framework that learns both solution rollouts and the scaffolds that guide them. DeepReinforce describes the core training idea as jointly optimizing task solutions and task-specific harnesses rather than relying on fixed human-written agent scaffolds, with the goal of discovering better search trajectories for software-engineering tasks. Public release materials present the family as spanning compact edge-deployable 9B models through frontier-scale 397B MoE variants, while the currently visible Hugging Face release set clearly includes 9B, 35B, and 397B checkpoints plus deployment-oriented FP8 and GGUF variants. The 35B model card describes reasoning-mode behavior with separate reasoning-content parsing, OpenAI-compatible tool calling, and deployment support across Transformers, vLLM, SGLang, llama.cpp, Ollama, and related local-serving stacks. Release benchmarks emphasize strong performance on agentic coding evaluations including Terminal-Bench 2.1, SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual, NL2Repo, ClawEval, and SWE Atlas, with the flagship 397B release positioned against top open models and even compared directly with Claude Opus 4.7 on Terminal-Bench 2.1 and SWE-Bench Verified.