DeepReinforce releases Ornith-1.5 open-source models with self-improvement loop

DeepReinforce launched the Ornith-1.5 open-source model series on August 19. The training uses a 'self-improvement loop' where the model creates harder tasks for itself to boost performance. The largest model, with 397B parameters, beats Claude Opus 4.8 in some tests.

The series has three sizes: Ornith-1.5-397B (MoE, 397B total) matches Claude Opus 4.8; Ornith-1.5-35B-A3B (MoE, 35B total) beats Muse-Glimmer-30B and Gemma-4-31B; Ornith-1.5-9B (dense) outperforms the larger Gemma-4-31B in most tests. There's also a quantized version, Ornith-1.5-9B-Mobile, that runs on Android phones and iPhone 17.