Cognition launches SWE-1.7 model with improved RL training techniques
HIGHLIGHTS
- •Cognition (Devin) launched SWE-1.7 model
- •Improved RL training techniques, infrastructure, training stability, high-quality data curation, and long-horizon task techniques
- •Achieves frontier-level intelligence at a lower cost, challenges the 'post-training ceiling' notion, and demonstrates the potential for RL to enhance model capabilities
Published:
Official Source (Official Blog)