Ornith-1.5: A New Family of Open-Source LLMs for Reasoning and Coding

Ornith-1.5 has been released as a new family of open-source LLMs, featuring a 9B Dense model, a 35B MoE model, and a 397B MoE model. These models are trained using self-improving strategies and achieve competitive performance on reasoning, agentic, and coding tasks. The release provides the open-source community with highly capable models that rival proprietary systems like Claude Opus 4.8 on complex benchmarks. It demonstrates the viability of self-improving training strategies in scaling model performance across different sizes. Ornith-1.5 achieves high scores on advanced benchmarks, including 56 on DeepSWE, 86.1 on Terminal-Bench 2.1, and 86 on SWE-Bench (verified). The models are hosted on Hugging Face under the Ornith-AI collection.