Ornith-1.0-35B Explained: Why It Beat A 397B Model

Описание: #Ornith1_0_35B #LLM #AI #OpenSource #AgenticCoding

Ornith-1.0-35B is a 35B-A3B MoE agentic coding model from DeepReinforce AI that scores 64.2 on Terminal-Bench 2.1 — a 22 point jump over its Qwen3.5-35B base. It matches or beats the Qwen3.5-397B on Terminal-Bench, SWE-Bench Pro, SWE-Bench Multilingual, and ClawEval, using roughly one tenth the active parameters per token. The recipe is self-scaffolding reinforcement learning: the model jointly optimizes the search scaffold and the solution it produces. MIT licensed. The community already shipped GGUF, FP8, NVFP4, and MLX ports in the first twelve days. We break down the numbers, the recipe, and what the Hugging Face discussion tab is asking for next.