Live drop
CognitionDropped Sep 10, 20264 builds4 makers1 prompt
Post-trained coding model on Kimi K3 base with unified RL for multiple reasoning effort levels at up to 70% lower cost.
About this release
SWE-2 is Cognition's most advanced coding model, post-trained from the 2.8T-parameter Kimi K3 open-weight base using scaled reinforcement learning. It introduces explicit reasoning effort levels trained in a single run and ships exclusively inside Devin Desktop, CLI, Web, and Fusion. It targets real-world software engineering workflows with strong results on FrontierCode, DeepSWE, and Terminal-Bench evals.
Read the official Cognition launch pageWorth opening first
Paste the URL and Dropday.ai will credit you, track the build as it evolves, and put it on this drop.
Submit a build