tl;dr: Cognition released SWE-1.7, an RL-post-trained model built on Kimi K2.7 that approaches GPT-5.5 and Claude Opus 4.8 performance on coding benchmarks (Terminal-Bench 2.1, SWE-Bench Multilingual, FrontierCode) at lower cost. Key technical contributions include top-p sampling replay to prevent entropy collapse, multi-cluster RL training across three continents using compressed weight deltas via object storage, self-compaction enabling six-hour rollouts, and rigorous data-quality pipelines to prevent reward hacking. The model shows notably more thorough codebase exploration and condensed chain-of-thought, and is available in Devin via Cerebras at 1000 TPS.