SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence (cognition.com)

Cognition has announced the launch of SWE-1.7, its most advanced AI model to date, which achieves frontier-level intelligence with significantly improved cost-effectiveness. This model boasts enhancements from its predecessor, Kimi K2.7, including stabilized training, refined infrastructure, and a high-quality data pipeline, ultimately pushing the boundaries of reinforcement learning (RL) capabilities. SWE-1.7 is particularly adept at long-horizon asynchronous tasks, critical for effective software engineering, which underscores its potential impact on the AI/ML community. Key innovations behind SWE-1.7 include multi-cluster training across three continents, addressing training instability through entropy preservation, and self-compaction techniques that extend task durations beyond traditional context limits. The model's architecture allows for efficient deployment of updates via weight delta transfer and fault-tolerant designs that minimize inference downtime during training disruptions. These advancements not only enhance model performance but also challenge existing assumptions about the limits of RL post-training, paving the way for future AI developments that harness large-scale, distributed compute resources effectively.