Ornith-1.5 Open Models Released in 397B, 35B, and 9B Sizes

Ornith has unveiled the Ornith-1.5 open model series. These models expand upon the previous version's self-scaffolding framework to implement a closed-loop self-improvement cycle, where the model proposes tasks, generates scaffolds, and derives solutions for reinforcement learning. The released lineup consists of a 397B MoE flagship, a 35B MoE model activating 3B parameters per token, and a 9B Dense model that includes quantized mobile builds for iPhone and Android.

The training cycle proceeds in three stages: proposing tasks that exceed the model's current capabilities, generating scaffolds to solve them, and producing solution rollouts. Reward signals are propagated across all three stages, enabling the simultaneous learning of better solutions, useful training tasks, and reliable evaluation harnesses. Task difficulty targets an empirical success rate of 0.2, and tasks that fail validity checks are assigned a reward of 0.

In benchmark results, Ornith-1.5-397B scored 85.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, demonstrating performance comparable to Claude Opus 4.8. The 35B model achieved 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified, while the 9B model outperformed Gemma 4-31B and Qwen 3.6-35B. Additionally, the flagship model recorded 92.8 on GPQA Diamond and 86.6 on BrowseComp, proving its strong performance in reasoning and agentic tasks.