Cognition's SWE-2

SWE-2 represents a significant advancement in coding models, delivering enhanced performance and cost-efficiency. Key capabilities include: * Superior code generation on benchmark tasks * Optimized for both capability and operational cost * Reduced turns required for task completion * Improved cost-performance ratio compared to predecessors This model is post-trained from Kimi K3, utilizing a novel reinforcement learning approach that simultaneously optimizes for diverse reasoning-effort levels. This method advances the entire cost-performance frontier, ensuring a highly competitive and efficient coding experience. SWE-2 achieves a 50.0% score on FrontierCode 1.1 Main, matching or exceeding other leading models in performance while offering substantial cost reductions. Compared to SWE-1.7, SWE-2 completes tasks in 58% fewer turns and costs 81% less, all while achieving a higher score. It also aligns closely with top-tier models like Fable 5.1 and GPT-6 Astra in capability but at a fraction of their cost. This makes it an ideal solution for developers and engineering teams seeking powerful, cost-effective code generation. SWE-2 is currently available through Devin Desktop and the command-line interface, providing flexible access for various development environments. It is perfectly suited for software engineers, research teams, and organizations focused on optimizing their development cycles and resource allocation.