Cognition launches SWE-2 AI coding model to reduce costs

Cognition has announced the launch of SWE-2, a new AI coding model designed to significantly reduce operational costs and increase efficiency in software development. The model is built upon Moonshot’s Kimi K3, which utilizes a Mixture-of-Experts architecture with 2.8 trillion parameters.

According to the manufacturer, SWE-2 reduces inference costs by 64% compared to previous versions. The model also demonstrates improved speed, requiring only 18 steps to reach its first code adjustment, a 62.5% reduction from the 48 steps required by its predecessor, SWE-1.

In industry benchmarks, SWE-2 has shown competitive performance. It achieved a 50.0% score in the FrontierCode 1.1 Main benchmark, outperforming models such as GPT-5 6 Sol and Grok 4.6. It also recorded a 92.8% success rate in the Terminal-Bench 2.1 test and a 73.0% rate in the DeepSWE-1.1 benchmark.