ornith-ai has released Ornith-1.5-35B-A3B, a mid-size mixture-of-experts model that activates roughly 3B parameters per token, on Hugging Face. The model extends Ornith-1.0 with a self-improvement loop that jointly optimizes task generation, scaffold construction, and solution rollouts via reinforcement learning. The card claims it outperforms Qwen 3.6-35B on coding and agentic benchmarks and exceeds dense models like Gemma 4-31B on agentic coding. Usage instructions cover Transformers, vLLM, SGLang, and Docker.
No score is assigned. Sources and their independence are shown in the citation chain below.