Ornith-1.5 open-source LLM family launches, rivals Claude Opus 4.8

Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, on par with Claude Opus 4.8 (85.0/59.0) and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731. Trained via end-to-end self-improvement — the model proposes tasks, generates scaffolds, and produces rollouts. The 9B-Mobile runs on phones; the 35B activates only 3B params per token.