DeepReinforce releases Ornith-1.5 in 397B, 35B and 9B sizes
TL;DR
- DeepReinforce released Ornith-1.5 in three sizes: a 397B MoE flagship, a 35B MoE activating 3B parameters per token, and a 9B dense model.
- The 397B scores 85.1 on Terminal-Bench 2.1, 56.0 on DeepSWE, 92.8 on GPQA Diamond, and 86.6 on BrowseComp, reported on par with Claude Opus 4.8.
- The training loop proposes its own tasks, builds scaffolds, and generates rollouts, optimized with GRPO and rewards tuned toward a 0.2 success rate.
DeepReinforce released Ornith-1.5 on Hugging Face in three sizes: a 397B mixture-of-experts flagship, a 35B MoE that activates 3B parameters per token, and a 9B dense model shipping with "a quantized Mobile build for iPhone and Android."
The 397B scores 85.1 on Terminal-Bench 2.1, 56.0 on DeepSWE, 92.8 on GPQA Diamond, and 86.6 on BrowseComp. TestingCatalog frames the terminal and coding numbers as "on par with Claude Opus 4.8 at 85.0 and 59.0" and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731. Neither the collection page nor TestingCatalog surfaces third-party reruns, so treat the parity read as the lab's own.
The training pipeline is the interesting part. The system runs three stages, all optimized with GRPO: it "proposes progressively harder tasks," generates task-specific scaffolds, and "produces the rollout conditioned on both." Rewards weight validity, novelty, and a frontier-difficulty term pinned to "targeting 0.2 success rate." The curriculum keeps the model at the edge of its own competence, without a human-curated task set.
Weights land under an MIT license, extending the pattern from Ornith-1.0 in June 2026. Two experts in our Who's Who directory shared the collection the day it dropped.
Shared on Bluesky by 2 AI experts
Originally reported by huggingface.co
Read the original article →Original headline: Ornith-1.5 - a ornith-ai Collection