Ornith-1.0: self-improving open-source models for agentic coding

	Ornith-1.0: self-improving open-source models for agentic coding(github.com)
	251 points by danboarder 1 day ago \| 48 comments
	tl;dr: Ornith-1.0 is an MIT-licensed family of open-source coding models (9B dense, 35B and 397B MoE) post-trained on Gemma 4 and Qwen 3.5, claiming SOTA results among open models on Terminal-Bench, SWE-Bench, and related agentic coding benchmarks. Its key trick is an RL framework that jointly optimizes both the solution rollouts and the scaffolds that generate them, allowing the model to discover better search trajectories. The checkpoints support 256K context, expose an OpenAI-compatible API via vLLM/SGLang, and work with agent harnesses like OpenHands, Claude Code, and OpenCode.
	HN Discussion: ↓Underwhelming real-world performance with hallucinations despite good benchmark scores ↑Genuinely useful Qwen fine-tune with creative coding solutions and faster inference ↓Skeptical this is just a benchmaxxed rebrand of existing Qwen/Gemma models ~Hardware requirements make models inaccessible to typical users •Questions about provenance, self-improvement mechanism, and missing model sizes