DeepReinforce releases open coding models that learn their own RL scaffolds

DeepReinforce has released Ornith-1.0, an open-source family of coding models built on Gemma 4 and Qwen 3.5. The project’s core idea is to let the model learn its own reinforcement-learning scaffold instead of using a fixed external harness.

The flagship 397B model reportedly reaches 82.4 on SWE-Bench Verified, with weights released under the MIT license. That puts the release squarely in the fast-moving race to make open coding models more capable on real software tasks.

For developers, the interesting part is not only the benchmark number but the training pattern: scaffolding itself is becoming a learned capability.