Kimi K3.1 Pro
Kimi K3.1 Pro is a multimodal release derived from K3. The model combines 2.8 trillion total parameters with 104 billion active parameters per token.
More than a new version number
The model card describes a change from K3’s Gated MLA layers to Gated DeepSeek Sparse Attention. It also lists quantization-aware training with MXFP4 weights and MXFP8 activations.Kimi K3.1 Pro model card
Practical guidance. When evaluating a local deployment, check support for this specific release in the inference engine. Support for K3 is not enough evidence that K3.1 Pro will run with the same configuration.
Keep the complete assistant message
The API requires the complete returned assistant message in later turns, including reasoning and tool calls. Vision requests use base64 data or uploaded-file references, rather than public image URLs.Kimi K3.1 Pro API guide
Practical guidance. A chat adapter that saves only the visible answer can lose required state. Check a task that calls a tool, receives its result and then continues.
Read the model-service conditions
The license requires a separate agreement for a model-service business whose combined revenue with affiliates exceeds $20 million over any consecutive 12 months. It also defines display requirements for certain large commercial products and exceptions for internal use and approved hosted access.Kimi K3.1 Pro license
Practical guidance. Check the exact license text against your intended distribution. Do this before committing to a self-hosted service for other customers.
Highspeed endpoint
Highspeed serves K3.1 Pro through a faster endpoint. Moonshot reports roughly two to three times the speed of the standard endpoint; this is a provider claim, not a guarantee for every request.Kimi K3.1 Pro API guide
Practical guidance. Use the exact endpoint identifier and price when comparing response times and costs. Compare complete tasks with the same prompts and tool inputs.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | kimi-k3.1-promoonshotai/Kimi-K3.1-Prokimi-k3.1-pro-highspeed |
| Context window | 1,048,576 tokens Published API context limit. Input, reasoning and output must fit within the request budget. Kimi inference billing |
| Attention design | KDA and Gated DSA; the latter replaces K3’s full-attention Gated MLA layers.Kimi K3.1 Pro model card |
| Reasoning | Always enabled; low, high or max effort. Max is the default.Kimi K3.1 Pro API guide |
| API entry requirement | A successful top-up of at least $1.Kimi K3.1 Pro API guide |
| Weight license | Kimi K3.1 Pro License.Kimi K3.1 Pro license |
Related models
- Kimi K3
Moonshot AI · Kimi K3
A long-context Kimi model with always-on reasoning and a million-token API context.
Launch date not confirmedAPI · Downloadable