Ornith Is the Open-Source Coding Model Built for Agents, Not Humans
DeepReinforce released Ornith-1.0, an MIT-licensed open-source family of coding models in four sizes: 9B, 31B, 35B MoE, and 397B MoE. It is designed for agentic coding, meaning it can plan, use tools, run tests, diagnose failures, and iterate with minimal human input. A key feature is that it learns its own task scaffold during reinforcement learning rather than using a fixed human-designed workflow. The lab says it added safeguards against reward hacking by keeping the environment and tests immutable, monitoring restricted actions, and using a frozen judge model as a veto layer. The 397B model scores 82.4 on SWE-bench Verified and 77.5 on Terminal Bench 2.1, ahead of Claude Opus 4.7 and DeepSeek-V4-Pro on those tests. On the harder SWE-bench Pro, it scores 62.2. The 9B model is notably strong for its size, scoring 69.4 on SWE-bench Verified. Ornith-1.0 is specialized for coding agents, not general-purpose chat or writing tasks.
