Most AI models are designed to chat with humans. Ornith doesn’t care about that. DeepReinforce’s new open-source model family is built for one thing: AI coding agents that work autonomously in real terminal and repository environments.
Released June 25 under the MIT license, Ornith-1.0 comes in four sizes — 9B, 31B, 35B mixture-of-experts, and a massive 397B MoE flagship. The 9B variant scores 69.4 on SWE-bench Verified, beating Google’s Gemma 4-31B, which manages 52.0. And the 9B model can run on a good smartphone.
What makes Ornith different? It doesn’t follow a human-designed playbook. Instead of relying on a fixed scaffold — rules for when to call a tool, how to handle errors — it develops its own strategies through reinforcement learning. The reward flows back to both the strategy and the code generation, so it learns how to plan, not just how to write.
DeepReinforce thought about reward hacking. If the model writes its own scaffold, it could theoretically game the verifier. Their fix: three layers of defense. The test suite is immutable and outside the model’s reach, a deterministic monitor flags restricted path access, and a frozen judge model evaluates results.
The model card is refreshingly honest — it may underperform on non-coding tasks. It’s wired for developer pipelines, not general conversation. That’s by design.
The parameter range is wide. The 9B dense model is accessible. The 397B MoE requires serious compute. All versions are MIT licensed with no regional restrictions, available on Hugging Face now.
