Researchers have launched Ornith-1.0, an open-source family of self-improving language models designed for agentic coding tasks. The lineup spans from compact 9B Dense variants for edge deployment to massive 397B Mo E models, built on Gemma 4 and Qwen 3.5 foundations. Ornith-1.0-397B achieves 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, matching Claude Opus 4.7 performance while outperforming similarly sized open models.

The key innovation is a self-improving training framework where the model jointly learns to solve tasks and construct the scaffolds guiding those solutions. Instead of relying on fixed human-designed harnesses, Ornith-1.0 generates both solution rollouts and task-specific harnesses simultaneously, creating a feedback loop that discovers better search trajectories. The system uses staged RL with staleness weighting to handle long rollouts and prevent reward hacking.

Performance benchmarks show Ornith-1.0-9B matching or exceeding larger models like Gemma 4-31B, demonstrating strong agentic coding capabilities in resource-efficient deployments. The 397B variant surpasses Mini Max M3 and Deep Seek-V4-Pro, while the 35B model outperforms much larger predecessors. This represents a significant advance in open-source coding assistants.