DeepReinforce releases Ornith-1.0, an open-weight coding model that writes its own RL scaffolds

DeepReinforce released Ornith-1.0 on June 25, an open-weight family of coding models under the MIT license whose central idea is self-scaffolding: the model generates its own reinforcement-learning training harnesses rather than relying on human-designed ones.

The technique matters because building RL harnesses by hand is one of the slowest parts of training coding agents. Instead of running on a fixed human-built harness, the Ornith-1.0 models generate both solution rollouts and the task-specific harnesses that guide them.

The family ships in four variants — a 9-billion-parameter dense model, a 31B dense model, and 35B and 397B mixture-of-experts models — built on a Gemma 4 and Qwen 3.5 foundation. The 35B model carries a 262,000-token context window and uses Grouped-Query Attention with SwiGLU.

DeepReinforce says its largest model outperforms a leading proprietary system on two coding benchmarks. The company reports that Ornith-1.0-397B scored 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, against figures it cites of 80.8 and 70.3 for Anthropic’s Claude Opus 4.7. Those scores are vendor-reported and not independently verified, and benchmark results from a model’s own developer routinely shrink under outside testing.

The release is fully open: all variants are available under the MIT license with no regional restrictions, a looser term than the custom licenses attached to many recent open-weight models.

Whether self-scaffolding holds up will turn on independent SWE-Bench runs and on whether developers keep the open-weight builds near the top of community download charts past the launch surge.