Moonshot AI releases Kimi K2.5

Moonshot AI released Kimi K2.5, an open-weight update to its Kimi K2 line built through continued pretraining of the existing 1-trillion-parameter, 32-billion-active-parameter mixture-of-experts model on roughly 15 trillion additional mixed visual and text tokens. The update was Moonshot’s first to give the model native image and video understanding, rather than bolting on a separate vision system, and it could reconstruct a working website layout directly from a video walkthrough.

Moonshot reported state-of-the-art open-weight scores across several benchmark categories, including 50.2% on Humanity’s Last Exam, 74.9% on BrowseComp, 78.5% on MMMU Pro and 76.8% on SWE-bench Verified — comparisons Moonshot ran itself rather than figures independently reproduced. The release also expanded Agent Swarm, the company’s orchestration layer for running many sub-agents on one task, to coordinate up to 100 parallel sub-agents across roughly 1,500 steps, a claimed 4.5 times speed-up over running tasks sequentially — though early users reported the swarm mode burned through tool-call quotas quickly.

The model was made available through Ollama, Together AI, Fireworks and OpenRouter, alongside a new Kimi Code subscription plan that drew some criticism for lower usage limits than competing coding-assistant subscriptions once an introductory pricing period ended. K2.5 was one entry in a rapid cadence of Kimi releases through 2026, followed by K2.6 in April, K2.7 in June and K3, a larger open-weight model, in July.