Cognition launches SWE-2 autonomous coding model
Cognition has introduced SWE-2, an autonomous coding foundation model post-trained from Moonshot AI's Kimi K3 using reinforcement learning across variable reasoning effort levels. Available within Devin Desktop and CLI, SWE-2 is engineered to optimize both long-horizon task completion and inference cost efficiency, scoring 50.0% on the FrontierCode 1.1 Main benchmark to rival leading proprietary models like Fable 5.1 while cutting inference expenses by up to 64%.
Cognition's shift toward custom domain post-training on capable open-weight foundations demonstrates that agent-specific reinforcement learning can outmaneuver general-purpose frontier models on real-world engineering tasks while drastically undercutting operating costs. Training across multiple reasoning effort tiers in a single RL run enables granular compute scaling, achieving near-frontier benchmark results at roughly one-third the inference expense. Embedding a specialized post-trained model directly into Devin Desktop and CLI protects Cognition from third-party model dependencies and API price volatility, while architectural optimizations reduce context drift and looping errors during complex codebase modifications.
DISCOVERED
59m ago
2026-09-11
PUBLISHED
1h ago
2026-09-11
RELEVANCE
AUTHOR
AICodeKing
