ornith 1.0

The Ornith team has released Ornith-1.0, a new family of open-source large language models purpose-built for agentic coding. Announced on June 25, 2026, the lineup spans multiple scales — a 9B dense model, 31B dense, 35B MoE, and a flagship 397B MoE — all post-trained on Gemma 4 and Qwen 3.5 bases. The standout technical contribution is a self-improving training strategy that uses reinforcement learning to jointly optimize not just solution rollouts but also the task-specific scaffolds that guide them, enabling higher-quality performance in complex, multi-step coding agent workflows. All models are released under the MIT license, with weights on Hugging Face, GGUF quantizations for local inference via Ollama and similar tools, and a supporting technical blog.Early evaluations show Ornith-1.0 achieving state-of-the-art results among open-source models of similar size on key agentic coding benchmarks. Highlights include 77.5 on Terminal-Bench 2.1, 82.4 on SWE-Bench Verified, strong multilingual SWE-Bench scores, and competitive results on NL2Repo, SWE Atlas, and ClawEval. Notably, the more efficient 35B MoE and 9B variants deliver best-in-class performance in their size classes, with the 35B even surpassing some much larger models on specific tasks. The release has been welcomed by the open-source community for combining strong capabilities with full commercial freedom and practical local deployability, adding momentum to the rapid progress in specialized coding agents.

models: https://huggingface.co/collections/deepreinforce-ai/ornith-10
blog: https://deep-reinforce.com/ornith_1_0.html