Ornith-1.0: MIT-licensed coding agents that learn their own scaffolds via RL

DeepReinforce AI has released Ornith-1.0, a family of open-source models tuned for agentic coding. The lineup spans four sizes — 9B dense, 31B dense, 35B MoE, and 397B MoE — built on top of Gemma 4 and Qwen 3.5 checkpoints. The team claims state-of-the-art results among open models in each size class across coding benchmarks including Terminal-Bench 2.1, SWE-Bench, NL2Repo, and ClawEval, with the smallest dense model fitting on a single 80GB GPU and the larger MoE checkpoints requiring tensor-parallel sharding across a node.

The notable technical claim is the training method. Rather than using reinforcement learning only to optimize solution rollouts, Ornith applies RL to jointly learn the scaffold — the search and tool-use loop that drives those rollouts — alongside the solutions themselves. The idea is that co-optimizing the agent’s control flow with its outputs lets the model discover better problem-solving trajectories instead of relying on a fixed, human-designed harness. The models are reasoning models that emit a chain-of-thought block before answering and produce structured tool calls.

Practically, every checkpoint exposes an OpenAI-compatible API with a 256K-token context window and native tool calling, served through recent builds of vLLM, SGLang, or Transformers. The MIT license and absence of regional restrictions are pitched as differentiators against more locked-down releases, making the weights usable in commercial agent frameworks and MCP-based tooling out of the box. As with any vendor-reported benchmark suite, the SOTA claims rely on the authors’ own harnesses and decoding settings and would benefit from independent reproduction.