Ornith ships four open coding models
DeepReinforce's Ornith 1.0, out June 25, is four MIT-licensed coding models from 9B to 397B that learn their own RL scaffold and post strong benchmark scores.
DeepReinforce released Ornith 1.0 on June 25, a family of four open coding models: a 9B and a 31B dense model, a 35B mixture-of-experts model, and a 397B mixture-of-experts flagship. All four are MIT-licensed and free to download, and all are post-trained on top of Gemma 4 and Qwen 3.5 base models.
The interesting part is how they were trained. Normally engineers wrap a model in a fixed harness, the part that decides which tool to call, when to retry, and how to break a job into steps, then tune that harness for each kind of task. Ornith folds that work into the model itself: during reinforcement learning it learns to produce both its answer and the harness guiding it. DeepReinforce calls the approach self-scaffolding.
On the lab’s own numbers, the 397B flagship reports 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, which would beat Claude Opus 4.7 and trail only Opus 4.8’s 87.6 on SWE-Bench. The small 9B model scores 69.4 and 43.1 on the same two tests and is built to run on edge devices. As with any vendor leaderboard, those figures are self-reported.
What it means for you
An openly licensed model you can run and modify, that lands near the top coding benchmarks, is a real option for teams that want to avoid a hosted API. It joins a fast-moving open field, from GLM-5.2 on the open-weights leaderboard to Kimi K2.7-Code. Before you trust the scores, run the verification habit on your own code, not the benchmark.