Kimi K3.1 Pro reaches hosted services; model weights are due by November 4

The launch notice separates hosted availability from the promised release of model weights by November 4, 2026. Kimi Work users need desktop version 3.1.0 or later, on Windows or an Apple silicon Mac.

For developers, the reasoning guide describes a model that always reasons. The optional reasoning_effort setting selects low, high or max effort; max is the default. This control balances reasoning depth and response time.

For conversations with several turns and for tool calls, the application must return the complete assistant message in the next request, including reasoning and tool-call fields. Keeping only the visible answer loses the required message structure.

The vision documentation supports image and video tasks. Send visual content as an array of message parts, not a string containing that array. Public image URLs are unsupported: use encoded image data or uploaded file identifiers. For a video, upload the file first, then reference its identifier. SVG drawings must be supplied as XML text rather than image input.

The billing guide distinguishes usage-based API charges from Kimi membership and Kimi Code. Both input and output are metered. File extraction and storage APIs are temporarily free, but extracted document content sent to the model still counts as billable input.

MarkTechPost also covered the launch.