Another brilliant open-source model release.

Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE,

Only 397B weights, but 4 wins over Claude Opus 4.8:

Terminal-Bench 2.1 (86.1 v 85), SWE-bench Verified (86 v 85.8), WideSearch (80.8 v 72.9), BrowseComp (86.6 v 84.3).

The architectural jump for Ornith-1.5 is that it effectively puts task generation, agent design, and solution generation inside the same reinforcement-learning loop. Improvement in one part can feed the other 2.

i.e. the model proposes its own tasks, writes the scaffold that will grade them, produces solution rollouts, and propagates reward back through all 3 stages with GRPO.

So the self-generated curriculum removes many human task design. Ornith begins with an environment or codebase plus high-level task instructions, then searches near the model’s current capability frontier.

That makes repositories and other verifiable environments more valuable training inputs than static examples alone. Labs with richer places to generate and verify new tasks could improve faster even without starting from the largest human-built dataset.