Kimi K3.1 Pro: hosted API, self-hosting and licence terms

Kimi K3.1 Pro is Moonshot AI's new open-weight multimodal model for agentic work, available both as downloadable weights and through Moonshot's API. For most teams the hosted API is the sensible way to trial it. Self-hosting needs serious GPU capacity, and the licence adds conditions for API resellers and very large consumer products.

What Moonshot AI released

Kimi K3.1 Pro was announced by Moonshot AI on 2 November 2026 with the model card in Moonshot's Hugging Face organisation. Moonshot describes it as a natively multimodal agentic model that takes text and images, post-trained from Kimi K3 with large-scale joint reinforcement learning, and the first model in the K3.1 series. The card names coding, 3D creation, financial research and factual accuracy as the areas that improved.

The facts that matter for planning, all from the model card:

• Architecture: a mixture-of-experts model with 2.8T total parameters and 104B active per token, 896 experts with 16 selected per token, plus a 401M-parameter vision encoder (MoonViT-V2).
• Context: a 1,048,576-token context length.
• Precision: MXFP4 weights and MXFP8 activations, trained with quantisation in mind rather than compressed afterwards.
• Access: the API model ID is kimi-k3.1-pro on platform.kimi.ai, with OpenAI-compatible and Anthropic-compatible endpoints. For self-hosting, Moonshot recommends vLLM, SGLang or TokenSpeed.

We could not read Moonshot's own blog post or API pricing page from our environment, so this post does not quote prices. Check them on platform.kimi.ai before you budget.

What changed from Kimi K3

The main structural change is in attention. According to the model card, the full-attention Gated MLA layers in Kimi K3 are replaced with Gated DeepSeek Sparse Attention layers for more efficient long-context modelling. The model now mixes 69 KDA layers with 24 Gated DSA layers.

Two behaviours are also worth knowing before you write code:

• Thinking effort: a top-level reasoning_effort field accepts low, high or max, and the default is max. If you send nothing, every call runs at the highest effort, which costs more and responds more slowly.
• Multi-turn and tool calls: Moonshot states that the complete assistant message must be passed back as-is, including reasoning_content and tool_calls, not just content.

The multi-turn rule that breaks existing agent code

That second point is the one most likely to cause quiet failures. Many agent frameworks and in-house wrappers built on the OpenAI chat format keep only content and tool_calls from each assistant turn and drop any extra fields. Pointed at Kimi K3.1 Pro through the OpenAI-compatible endpoint, that code will still run, but the model loses its earlier reasoning between turns.

Our advice: before trusting any evaluation, log the exact message list your framework sends on the second and third turns and confirm reasoning_content survives. If you route requests through a gateway or proxy, check that it does not strip unknown fields. Set reasoning_effort explicitly per task type, for example low for classification and routing steps and high or max only for planning and code generation. Our guide to context engineering for AI agents covers how to budget what goes back into each turn.

Self-hosting: hardware and licence checks

The open weights are the reason many teams will look at this model, so be clear about what they take. The files in the Hugging Face repository total about 1.56 TB even in MXFP4. That is a multi-node GPU deployment, not something a single inference server will hold, and a full-length context adds memory on top. Our post on running an AI agent on a self-hosted model walks through the operational side.

The Kimi K3.1 Pro licence is permissive in most respects: use, modification, fine-tuning and redistribution are allowed with the copyright notice kept. Two conditions apply on top:

• Model as a Service: if your company sells model inference or fine-tuning to third parties who control the inputs or parameters, and your group revenue passes the threshold the licence sets, you need a separate agreement with Moonshot before commercial use. Products that only embed the model inside their own features are excluded from this definition.
• Attribution for very large products: commercial products above the licence's monthly active user or monthly revenue thresholds must display "Kimi K3.1-Pro" prominently in the interface.

Neither condition applies to internal use that does not expose the model or its outputs to third parties, or to access through Moonshot's own products and certified inference partners. For most B2B agent products neither threshold will bite, but write the licence check into your model approval record. Our checklist for open-weight model licences lists what to record.

What it means for teams building agents

Kimi K3.1 Pro gives teams another open-weight option with a very long context and published agentic evaluations, alongside an API that can drop into existing OpenAI or Anthropic client code. The useful questions are practical ones:

• Benchmarks: Moonshot's own coding scores were mostly run inside the Claude Code harness, according to the evaluation notes on the model card. Your agent's harness, tools and prompts are different, so run your own task set before drawing conclusions.
• Data residency: the hosted API sends prompts to Moonshot. If your contracts restrict where customer data is processed, confirm where Moonshot and its certified partners run inference, or plan for self-hosting.
• Cost and latency: with max as the default effort, early tests can look slow and expensive. Compare like for like by fixing the effort level across every model you test.

When not to switch: if your current model meets quality targets and your data agreements do not cover a new provider, the gain may not be worth a new vendor review. And if you would self-host only to save money, price the GPU fleet against your actual traffic first; at low volume the API is usually cheaper.

Frequently asked questions

Is Kimi K3.1 Pro open source?

The weights are published on Hugging Face under the Kimi K3.1 Pro licence, which allows use, modification, fine-tuning and redistribution. It is not a standard open-source licence: companies selling model inference above a revenue threshold need a separate agreement, and very large consumer products must show the model name. Internal use is exempt from both conditions.

How do I call Kimi K3.1 Pro from existing code?

Moonshot provides OpenAI-compatible and Anthropic-compatible APIs on platform.kimi.ai, using the model ID kimi-k3.1-pro. Existing client libraries should work with a changed base URL and key. The important change is to pass the full assistant message back on every turn, including reasoning_content and tool_calls, and to set reasoning_effort explicitly rather than relying on the max default.

What hardware does self-hosting Kimi K3.1 Pro need?

The repository files total about 1.56 TB in MXFP4 precision, so expect a multi-node GPU cluster rather than a single server. Moonshot recommends vLLM, SGLang or TokenSpeed as inference engines and links recipes for each. Long contexts add memory for the cache on top of the weights, so size the cluster for the context lengths your agents actually use.

Should we replace our current model with Kimi K3.1 Pro?

Only after testing it on your own tasks in your own harness. The published coding results were mostly produced with Claude Code, which may not match how your agent works. Run a fixed evaluation set at a fixed effort level, check data processing terms with your legal team, and keep your current model as a fallback during any trial.