Cognition has announced the release of SWE-2, the company's most advanced coding model to date. According to the company, SWE-2 recorded a score of 50.0% on FrontierCode 1.1 Main, closely approaching the performance of Fable 5.1 while reducing costs by 64%. Furthermore, compared to GPT-6 Astra, it keeps the score difference to only a few points while reducing costs to one-quarter.

The development of SWE-2 utilized post-training techniques based on Kimi K3, which features 2.8 trillion parameters. A key characteristic is the introduction of a new algorithm in reinforcement learning (RL) that allows for the training of various effort levels in a single execution. The company explains that this achieves a Pareto optimal balance between cost and performance depending on the complexity of the task.

In terms of performance, SWE-2 medium recorded higher scores than the previous SWE-1.7 while reducing the average number of execution steps by 58%. This is attributed to improved model judgment, which reduces unnecessary exploration of the codebase and allows for faster implementation.

SWE-2 is available starting today on Devin Desktop, CLI, Devin Web, and Fusion.

Source: