Ornith-1.0: self-improving open-source models for agentic coding
Points and comments are a snapshot, not live.
Ornith-1.0 open-source models claim state-of-the-art coding agent benchmarks via a self-improving RL training framework.
Built on top of Gemma 4 and Qwen 3.5, Ornith-1.0 comes in 9B dense, 35B MoE, and 397B MoE variants. It employs reinforcement learning that jointly optimizes solution rollouts and the scaffold that drives them. On Terminal-Bench 2.1 (Terminus-2), the 9B scores 43.1, 35B scores 64.2, and 397B scores 77.5. SWE-bench Verified results are 69.4, 75.6, and 82.4 respectively. The models use 256K context windows and are MIT licensed. Quickstart guides include vLLM, SGLang, Hugging Face Transformers, and integration with frameworks like OpenHands and Hermes Agent.
What commenters are saying
Many commenters dismiss the 'self-improving' label as misleading, noting it refers to the training process (RL on generated scaffolds and solutions) rather than a model that improves at inference time. Several call it a fine-tuned, benchmaxxed version of Qwen and Gemma 4. Some users report positive experiences: one finds Ornith-1.0 35B slightly better than Qwen 3.6 35B on C++ codebase tasks with smaller CoT and faster output, while another flags poor hallucination in chat without tools. Others question missing 31B models and benchmark inconsistencies.