Ornith-1.0 drops with self-improving scaffolding

Ornith-1.0 is a major breakthrough for open-source agents, proving that reinforcement learning can optimize both task execution and execution scaffolds.

• –Direct training on scaffold generation reduces developer effort in manually designing complex execution environments
• –Built-in reward-hacking mitigation solves a critical issue where RL agents write passing tests instead of correct code
• –Broad range of sizes, from 9B dense to a 397B MoE flagship, enables deployments from local edge dev environments to massive enterprise clusters
• –Outperforming peers on SWE-Bench Verified (82.4) demonstrates that self-generated scaffolding matches or exceeds hand-crafted developer loops