Ornith-1.0: Self-scaffolding LLMs for agentic coding
Summary
Ornith-1.0 introduces a self-improving family of open-source LLMs for agentic coding, ranging from edge-friendly 9B dense models to 397B MoE frontier-scale models. It uses a self-improving training loop that co-evolves the scaffold and the solution, with benchmarks showing strong competitive performance versus larger proprietary models. The approach includes safeguards against reward hacking and describes an asynchronous RL training pipeline.