Kimi K3.1 Pro: Capabilities, Hardware, and Access Options

Kimi K3.1 Pro is Moonshot AI's post-trained successor to Kimi K3. It keeps K3's scale, 2.8 trillion total parameters with 104 billion active per token, and its 1,048,576-token upstream context. It changes how the model attends to long inputs, how much reinforcement learning shaped it, and one setting you'll notice quickly: thinking is always on.

We offer Kimi K3.1 Pro in Maple Chat and Maple Work on Pro, Max and Team plans, and through Maple Proxy, with a 262,144-token window. This guide covers what changed from K3, what self-hosting takes, and which route fits your work. If you already use K3, our Kimi K3 guide covers the model it builds on.

What is Kimi K3.1 Pro?

It is a mixture-of-experts model: 896 routed experts, of which 16 run for each token, plus two shared experts. That is why only 104B of its 2.8T parameters are active per token. Moonshot's card describes the release this way:

Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, it is the first model in the K3.1 series.

Active parameters describe compute per token, not memory. Every token can route to different experts, so all 2.8T must stay available to the serving system. Moonshot ships the routed experts in MXFP4, a 4-bit format the model was trained to use, which keeps the download near 1.56 TB.¹

The card lists text and image as its modalities, with MoonViT-V2, a 401M-parameter vision encoder. Some summaries of the launch list video input too. Moonshot's own API accepts video files,⁵ but the model card doesn't list video, and neither does our route.

Is Kimi K3.1 Pro open source?

Not in the standard sense: Moonshot publishes open weights under the Kimi K3.1 Pro License, a custom license rather than MIT or another standard one. We compared it with the Kimi K3 License, and apart from the model name the text is the same.

The license grants broad rights to use, modify and distribute the weights. It adds conditions for very large users. If you or an affiliate run a model-as-a-service business and your combined revenue tops $20 million over any 12 months, you need a separate agreement with Moonshot before commercial use.

A commercial product with more than 100 million monthly active users or $20 million in monthly revenue must prominently display the model's name on its user interface. The license gives that name as "Kimi K3.1-Pro".²

Neither condition applies to internal use or to use through Moonshot's own products and the inference partners it certifies. If you plan to build on the weights, read the license yourself and get legal advice where it matters.

What changed from Kimi K3?

The scale didn't. The post-training, one part of the attention design and the default behavior did. Moonshot calls K3.1 Pro "a post-trained version of Kimi K3, with substantially larger post-training compute."⁶

The architectural change is narrow. K3 mixes 69 Kimi Delta Attention layers with 24 full-attention layers; K3.1 Pro replaces those 24 with gated sparse attention, which Moonshot says is for "more efficient long-context modeling."^{1, 7} The routed experts, layer count and quantization scheme are unchanged.

Moonshot's quickstart says K3.1 Pro "always has thinking mode enabled," with low, high and max effort levels and max as the default, and it fixes temperature at 1.0 and top_p at 0.95.⁵ We follow those settings in Maple, which is why the Off option you may know from K3 isn't there.

Like K3, K3.1 Pro expects its earlier reasoning to be passed back on every turn. We do that. In long chats, that earlier thinking counts as input again on later turns, so the level you pick matters more than it seems.

What do Moonshot's evaluations show?

They show a model Moonshot places near the strongest proprietary models on coding and agent work, and behind them on several reasoning and security tests. These are Moonshot's results, run at max effort with its chosen harnesses, often the Claude Code or Kimi Code agent frameworks.¹

Against K3, Moonshot reports its internal long-horizon coding score rising from 55.95% to 61.2%, and the claim-level hallucination rate on its internal stress test falling from 23.4% to 8.3%.⁶ Moonshot qualifies the second number itself:

The reported rates do not represent overall error rates in daily use.

Moonshot also says it removed cyber-related training data before the open-weight release, and that K3.1 Pro's offensive security capability is below what its pipeline would otherwise produce.

A few habits help when reading all of this:

• Check the harness and effort level. Agent benchmarks depend heavily on the scaffolding around the model.
• Check which competitor versions were tested. These comparisons date quickly.
• Treat benchmarks as a shortlist signal. The test that matters is your own work at the thinking level you'll use.

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

Roughly what K3 needed. We checked both repositories' file listings on Hugging Face: each holds 96 weight files, and K3.1 Pro's total about 1.561 TB, only 357 MB more than K3's.⁸ The extra space is consistent with the sparse-attention indexer added to 24 layers.

If you already planned a K3 deployment, the weight budget carries over almost unchanged. You can reproduce the check with one request to Hugging Face's file-listing API for each repository and a sum of the file sizes.

vLLM's reference setup uses eight data-center GPUs in one tensor-parallel group: two NVIDIA GB300 trays or one B300 node, with throughput not yet tuned.³ On AMD, the recipe reports one MI355X node running a separate, tuned eight-GPU configuration.

With an FP8 cache on eight B300 GPUs, the recipe reports room for about 2.6 million cached tokens at the baseline, and about 18.5 million with eight-way decode context parallelism (DCP8), at the cost of somewhat slower decode steps. That is cache capacity shared across requests, not a longer context window for one request.

There is no Mac or consumer-GPU path in the recipe, and we don't offer a universal minimum either. Your needs depend on the quantization, the context you serve, concurrency and the serving framework.

Self-hosting is the right call in some cases:

• your information must stay on hardware you control
• you need to fine-tune, modify or deeply instrument the model
• you already run multi-GPU inference with capacity to spare

If none of those apply, the operational work may not buy you anything a hosted route can't.

How do you use Kimi K3.1 Pro in Maple?

Choose Kimi K3.1 Pro from the model picker in Maple Chat or Maple Work, our desktop app, on a Pro, Max or Team plan. It isn't on Free. Plan details are on our pricing page, and the model sits with the others we offer in our model library.

We serve it with a 262,144-token window, about a quarter of the upstream maximum and still enough for long documents, large parts of a codebase or a long conversation. Thinking runs at Low, High or Max, and starts at Max.⁴ Try Low for everyday questions, since it uses the least reasoning effort, and keep Max for hard coding, analysis and multi-step agent work. Our model guide explains how thinking levels affect usage.

You can upload an image while using Kimi K3.1 Pro in Maple Chat. The app first has a separate vision model describe it, then gives that description to the model you chose.⁹ The answer is based on the description, not a direct view of the image, which matters when a small visual detail decides the task.

For developer tools and your own applications, Maple Proxy is our API route, with the model ID kimi-k3-1-pro. It's billed per token at these rates, shown with Kimi K3's for comparison:¹⁰

USD equivalent per 1 million tokens. API usage is charged by the chosen model’s input and output tokens. It draws from your plan’s usage first, then from any extra credits you’ve bought. Cached-input rates apply only to tokens reported as cached. — means the standard input price applies. Automatic model aliases are billed at the rate of the model selected for the request.

Check the Proxy documentation for current input types and limits.

This route gives you two things. First, you use the model without buying, running or updating GPU servers like the eight-GPU setup in vLLM's recipe. Second, your requests go through our protected inference path, not a conventional hosted API; the next section explains it. The limits to plan around:

• Context is 262,144 tokens, not 1M. For more in one request, use Moonshot's API or your own deployment. In Maple, DeepSeek V4.1 Flash has the full 1,048,576-token window.
• No video input. Our route takes text and images.
• Thinking can't be turned off. Low is the lightest setting.

What information reaches each route?

That depends on who operates the model: you, a conventional host, or Maple. Several questions hide inside "is it private": who can read your prompts while they're processed, what happens to them afterward, and whether any of it trains a model.

Self-hosted weights. Prompts and outputs can stay on infrastructure you operate, as long as your tools and connected services don't send them elsewhere. The trust boundary is yours, along with the job of securing it.

A conventional hosted API. With Moonshot's API or another host, your requests are processed in a form the provider's systems can read. Retention, training use and staff access then depend on that provider's terms, which vary by service and plan.

Maple. Your prompts are encrypted before they leave your device and processed inside a secure enclave, an isolated computing environment. The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt, much as certificate chains work on the web. Our proof page describes the architecture and the attestation you can check yourself.

That design protects the contents of your chats: your prompts and the responses. Our privacy notice covers everything outside them. It also doesn't vouch for what the model's output leads you to do. If an agent runs K3.1 Pro's code or acts on its suggestions, review that work as you would any untrusted input.

Which route fits your task?

It depends on where your information may go, how much context you need and how much infrastructure you can run.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware or you need control over the weights.
• Choose Moonshot's API or another host if you need the full million-token window or video, and a conventional provider boundary fits your data.
• Choose Maple if you want K3.1 Pro on sensitive work without running infrastructure, and the work fits in 262,144 tokens of text and images, with images in Maple Chat passed as descriptions.

If you're moving from K3, run the same prompts on both at the thinking level you plan to use. The weights barely changed in size; the answers are where the difference shows.

Sources

1. Moonshot AI, Kimi K3.1 Pro model card. Parameters, experts, attention layers, quantization, modalities, context length, evaluation results and settings, preserved reasoning.
2. Moonshot AI, Kimi K3.1 Pro License. The custom license and its conditions for large services.
3. vLLM, Kimi-K3.1-Pro recipe. Baseline hardware layouts and cache capacity on eight B300 GPUs.
4. Maple, public model configuration. Plan tier, context window, thinking levels, sampling and inputs for Kimi K3.1 Pro and Kimi K3 in Maple.
5. Moonshot AI, Kimi K3.1 Pro API quickstart. Always-on thinking, effort levels, fixed sampling, 1M context and video input on Moonshot's API.
6. Moonshot AI, Kimi K3.1 Pro tech blog. Post-training, comparisons with Kimi K3, the hallucination caveat and the cyber-data removal.
7. Moonshot AI, Kimi K3 model card. Kimi K3's attention-layer composition.
8. Hugging Face, file listings for moonshotai/Kimi-K3.1-Pro and moonshotai/Kimi-K3. Number and size of weight files in each repository.
9. Maple, Maple Chat image handling and vision helper. Image description before the selected model responds.
10. Maple, pricing page, API prices per million tokens. Maple Proxy prices for Kimi K3.1 Pro and Kimi K3.