Self-improving open-source models emerge

Ornith-1.5 (9B-397B) introduces a self-improvement training paradigm where models iteratively improve their own performance, claiming SOTA and parity with Claude Opus 4.8. This approach could reduce reliance on massive datasets and human annotation, aligning with community goals of efficient fine-tuning and democratization. The model is open-source, with sizes ranging from 9B to 397B. Verification of release and license is pending, but the concept is a strong signal for the future of open-weight AI.