OSS
Ornith-1.5 Ships 9B Dense, 35B MoE and 397B MoE Under MIT, With the 397B Claiming 86.0 SWE-bench Verified Against Claude Opus 4.8's 85.8
DeepReinforce released the Ornith-1.5 family on August 19 under MIT in three sizes, and it is now the highest-trending non-Qwen text model on Hugging Face at a trending score of 424 for the 35B. The 397B flagship reports 86.1 Terminal-Bench 2.1, 86.0 SWE-bench Verified, 65.1 SWE-Bench Pro and 79.6 SWE-Bench Multilingual; the 35B activates only 3B parameters per token and reports 79% SWE-bench Verified and 89.2 GPQA Diamond. The training pitch is a closed self-scaffolding loop where the model proposes its own tasks, writes the scaffolds and generates its own RL rollouts, which is the part worth skepticism until someone reproduces it.
Source
↳ Follow the thread