Moonshot AI released the model weights and technical report for Kimi K3, a 2.8T-parameter mixture-of-experts model with native visual understanding and a 1M-token context window. The company said the new architecture delivers 2.5x the intelligence per unit of compute and said it is also open-sourcing more of the software stack behind the model, including attention kernels, a communication library for distributed MoE workloads and infrastructure for running agent environments at scale.

Kimi K3 was made available on Nebius on launch day, with day-0 support from vLLM and availability in Cognition's Devin Desktop and CLI. The Kimi K3 license says model-as-a-service businesses making more than $20M a year need a separate agreement, while products with more than 100M users or $20M a month in revenue must display "Kimi K3." Later in the day, Arena ranked Kimi K3 No. 1 among open-weight models in Agent Arena with +9.75% net improvement, ahead of GLM-5.2 Max at +7.12%.