China’s Kimi K3 AI Outperforms GPT 5.6 Sol in SWE Marathon

China’s Kimi K3 AI is presented as a 2.8-trillion-parameter open-access large language model.

It reportedly uses Kimmy Delta Attention and attention residuals to improve memory efficiency and output quality.

The article claims Kimi K3 can generate operating systems and game environments, and outperforms GPT 5.6 Sol on SWE Marathon tests.

If accurate, it could raise the bar for coding and agentic AI while making advanced models more accessible and cheaper to deploy.