DeepReinforce released Ornith-1.0, a 397B-parameter MoE coding model that hits 79.2% on SWE-Bench Verified — approaching Claude Opus 4.7 (81.5%). The most innovative thing: the model uses an RL harness that is partly written by the model itself.
The "self-written harness" mechanism: traditional RL training requires a hand-written reward function and a hand-written evaluation harness. Ornith-1.0's training pipeline has the model generate candidate harnesses (reward function + test cases + evaluation logic), and then uses a meta-verifier to select the most "honest" harness — one that scores real code quality rather than gaming the metric.
The training pipeline has three stages: SFT on public code → RL with the auto-generated harness → self-distillation. Each stage has a "self-eval" step — the model evaluates its own training progress and dynamically adjusts the learning rate, the reward function, and the data sampling strategy.
The result: on SWE-Bench Verified, Ornith-1.0-397B-MoE (50B active) hits 79.2% — significantly above Qwen3-Coder-480B (76.1%) and approaching Claude Opus 4.7. The model is fully open-sourced, including the training code, the data, and the auto-harness.
The bigger takeaway: "self-written harness" is a major step toward "self-improving AI." Ornith-1.0 demonstrates that a model can be trusted to design part of its own training loop — as long as there's a meta-verifier to catch the gaming. This is a significant relaxation of the "human-in-the-loop" assumption, and may become a key technique for the next generation of foundation models.