DeepReinforce just open-sourced Ornith-1.0, a family of agentic coding models running from a 9B Dense for edge boxes up to a 397B MoE. Built on Gemma 4 and Qwen 3.5, MIT-licensed, all weights already on Hugging Face. It hit the HN front page fast.
The model that builds its own scaffold
Everyone else hand-writes the harness that wraps a coding model — tool loops, prompts, the whole rig. Ornith treats that scaffold as something to learn. Each RL step, the model first writes a task-specific scaffold, then generates the solution against it. Model teaches itself how to set up the job before doing the job.
Numbers back it up: 397B scores 82.4 on SWE-Bench Verified, essentially level with Claude Opus 4.7 and ahead of MiniMax M3 and DeepSeek-V4-Pro. The tiny 9B still pulls 69.4 SWE-bench and 43.1 Terminal-Bench 2.1.
Run it yourself
No API to sign up for — grab deepreinforce-ai/Ornith-1.0-9B (or the 397B) off Hugging Face and deploy local. The 9B is small enough to put a self-scaffolding agent on your own GPU.
You Might Also Like
- Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
- Claude Code Remote Control Just Turned my Phone Into a Coding Terminal and im Weirdly Into it
- Minimax m2 7 Scores 56 on swe pro and it Helped Build Itself
- Qwen 3 6 Plus vs Claude Opus 4 6 3x the Speed 1 17th the Price and the Benchmarks are Uncomfortably Close
- Z ais glm 5 1 Tops swe Bench pro at 58 4 Trained on Zero Nvidia Hardware
Top AI Product 
Leave a comment