Agentic coding earns its family under the MIT license.

Deep Reinforce releases Ornith-1.0, a family of agentic coding LLMs from 9B to 397B post-trained from Gemma 4 and Qwen 3.5 under the open-source MIT license.

With Ornith-1.0, Deep Reinforce offers a family of LLMs specialized in agentic coding, ranging from 9B Dense to 397B MoE, including 31B Dense and 35B MoE versions.

The models are post-trained from Gemma 4 and Qwen 3.5, using a method called self-scaffolding: during reinforcement training, the model not only learns to produce a solution but also to generate the execution framework that helps it achieve it. The idea is to allow code-specific strategies to emerge, rather than relying solely on hand-designed structures. Ornith highlights strong results on Terminal-Bench 2.1, SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual, NL2Repo, SWE Atlas, and ClawEval, with favorable comparisons to other open models of similar size.

The team also describes several safeguards against reward hacking, including fixed environment limits, deterministic control over forbidden actions, and a frozen LLM judge layer. All models are released under the MIT license, allowing commercial use and research.