Ornith-1.0: self-improving open-source models for agentic coding
Summary
Ornith-1.0 introduces self-improving open-source coding agents in multiple sizes (9B, 35B, 397B) with an RL-based training loop that optimizes both the scaffolding and the solutions. It features an OpenAI-compatible interface and tool-calling, enabling agent frameworks and local deployment via vLLM, SGLang, or GGUF. Benchmarks show strong performance against size-matched baselines on coding tasks.