Kimi K3.1 Pro: architecture, benchmarks and licence

Kimi K3.1 Pro is an open-weight language model from Moonshot AI. It is a Mixture-of-Experts model with 2.8 trillion total parameters, of which 104 billion are active for each token. It reads text and images and handles up to 1 million tokens of context.

Moonshot describes it as a post-trained version of Kimi K3 and promised the full weights by 4 November 2026. They are now on Hugging Face. We read the model card, the Moonshot blog post and the licence. Below we separate what is confirmed, what Moonshot claims and what it has not disclosed.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is the first model in Moonshot's K3.1 series. Moonshot built it from Kimi K3 with "large-scale joint reinforcement learning across hundreds of downstream domains". It says the post-training compute was "substantially larger" than for K3. A faster variant, K3.1 Pro Highspeed, is offered through the API, and Moonshot says it runs 2 to 3 times faster.

You can use the model in the Kimi website and apps, the Kimi Work desktop app, Kimi Code and the API. Moonshot says the API accepts both OpenAI-style and Anthropic-style requests.

How big is Kimi K3.1 Pro, and how is it built?

The model card lists these facts:

• Size: 2.8T total parameters, 104B active per token.
• Layers: 93, of which 1 is a dense layer.
• Experts: 896 experts, 16 picked per token, plus 2 shared experts.
• Attention: 69 Kimi Delta Attention (KDA) layers and 24 Gated DeepSeek Sparse Attention (DSA) layers.
• Context: 1,048,576 tokens, about 1 million.
• Vocabulary: 160K tokens.
• Vision: a MoonViT-V2 image encoder with 401M parameters.
• Precision: MXFP4 weights and MXFP8 activations, learned through quantisation-aware training.

Look at the ratios first. Only about 3.7% of the parameters work on any one token (104B out of 2.8T). The router picks 16 of 896 experts, under 2%. This is how Mixture-of-Experts models stay affordable at huge sizes: you pay to store every expert but compute with only a few. Our post on small language models explains why active parameters drive cost.

MXFP4 is a 4-bit number format in which each small block of weights shares one scale factor. Training with it from the start, called quantisation-aware training, means the model learns to work at low precision instead of being squeezed after training. Smaller numbers mean less memory and faster serving.

What is hybrid attention in Kimi K3.1 Pro?

Kimi K3.1 Pro mixes two kinds of attention layer, about three cheap layers for every precise one. This mix is what makes a 1M-token context practical.

Standard attention compares every token with every earlier token. Its memory grows with the length of the text, which becomes very expensive at a million tokens. The two layer types deal with this in different ways:

• KDA layers: Kimi Delta Attention is a form of linear attention from Moonshot's Kimi Linear paper, published in October 2025. It keeps a fixed-size memory that it updates token by token, with its own forget rate for each feature channel. That memory stays the same size however long the text gets.
• Gated DSA layers: DeepSeek Sparse Attention uses a small, fast "indexer" to score all earlier tokens, then attends only to the top 2,048 of them. The model card lists exactly this setting.

Linear attention is cheap, but it can blur details across long texts. The sparse layers let the model look back precisely at the tokens that matter. The 69 to 24 split is close to the 3:1 ratio that the Kimi Linear paper used.

This is the one architecture change from K3 that Moonshot states. Kimi K3 used full-attention Gated MLA layers in those 24 places. For K3.1, Moonshot replaced them with Gated DSA "for more efficient long-context modeling".

How good is Kimi K3.1 Pro? Benchmarks as Moonshot reports them

All the numbers below are Moonshot's own, with reasoning effort set to max. Moonshot ran the closed models itself and notes that its testing of them "may involve some fallback behavior". Several benchmarks are internal and cannot be checked by anyone else. Read them as claims.

Kimi K3.1 Pro leads or comes close on several agent tasks that use tools:

• HLE-Full with tools: 70.0, against 67.7 for Claude Opus 5.5 and 57.2 for GPT-6 Astra.
• BrowseComp: 92.1, against 91.5 for GPT-6 Astra and 90.8 for Claude Opus 5.5.
• SWE Marathon v1.1: 51.7, the top score in Moonshot's table.
• Terminal-Bench 4.0: 65.7, close to Claude Opus 5.5 at 66.4.

It trails the best closed models on the hardest reasoning without tools, and on some coding tasks:

• HLE-Full, no tools: 51.1, against 59.8 for Claude Opus 5.5.
• HLE-Diamond: 39.4, against 65.8 for GPT-6 Astra.
• Terminal-Bench-Science 0.1: 38.1, against 68.1 for GPT-6 Astra.
• Program Bench: 31.5, against 65.0 for Claude Opus 5.5.

Our reading: the pattern fits a model trained hard on tool use. With search and code tools, it matches the frontier. On the hardest closed-book questions, a large gap remains. Since most people use these models with tools, the first list probably matters more in practice. To understand what the effort setting changes, read our guide to reasoning models.

Moonshot also reports fewer made-up facts on its internal stress test with tools: a claim-level rate of 8.3%, down from 23.4% for Kimi K3. It adds a fair warning: "The reported rates do not represent overall error rates in daily use."

Did Moonshot remove cyber skills before release?

Partly, and this is one of the more unusual parts of the release. Moonshot says cyber-related data was in the mid-training data, and cyber tasks were in the reinforcement learning mix. For the open-weight release, it "identified and removed cyber-related data from both training stages".

Moonshot gives one number for the effect. On its internal xdev-bench-hard test, a version trained without these measures scored 90.5% pass@3. The released model scores 81.0%. Moonshot says the model's cyber ability "remains behind that of frontier proprietary models that have been publicly available for at least three months."

This is a real research problem for open models. Once weights are public, nobody can add safety filters later, so any reduction in risk has to happen during training. A drop from 90.5% to 81.0% is a reduction, not a removal. For why this matters, see our post on how OpenAI's agents found 948 Linux root exploits.

What is the Kimi K3.1 Pro licence?

The Kimi K3.1-Pro License starts like the MIT licence. You may use, copy, modify, share and sell the weights and code for free, as long as you keep the copyright notice. It then adds two conditions for very large businesses:

• Hosting services: if you sell access to the model as a service and your revenue passes 20 million US dollars in any 12 months, you must sign a separate agreement with Moonshot first.
• Big products: a commercial product with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, must show "Kimi K3.1-Pro" prominently in its interface.

The licence exempts internal use that does not make the model available to others. Below those limits the extra conditions do not apply, so for students and small teams it works like an open licence. It is still a custom licence, not a standard one such as Apache 2.0, so read it before building a large commercial service.

How to use Kimi K3.1 Pro

There are three ways to use Kimi K3.1 Pro, from easiest to hardest:

1. 1Use the app: the model is in the Kimi website, the mobile apps and the Kimi Work desktop app. Enterprise accounts need an admin to switch it on.
2. 2Use the API: the API at platform.kimi.ai accepts OpenAI-style and Anthropic-style requests. Set reasoning_effort to low, high or max; the default is max. In multi-turn tool use, send the whole assistant message back, including reasoning_content and tool_calls, not just the text. Check current prices on the platform.
3. 3Run it yourself: the weights are on Hugging Face. Moonshot recommends vLLM, SGLang or TokenSpeed for serving.

Moonshot does not list hardware needs. Our rough estimate: 2.8 trillion weights at 4 bits each take about 1.4 TB, before any memory for the context. That means several multi-GPU servers, not a laptop. For experiments on your own machine, a small open model is the practical choice. See our guide to running Gemma 4 on a laptop.

What has Moonshot not disclosed?

There is no technical report for Kimi K3.1 Pro. The model card and blog do not state:

• how many tokens the model was pretrained on, or the data mix
• the optimiser, training hardware or training time
• how much reinforcement learning compute was used, or how many environments
• hardware needs and API prices

The architecture table is unusually detailed for a model this strong, so researchers can still study the design. But claims about why the model performs well cannot be checked without the training details.

Research ideas for students using Kimi K3.1 Pro

An open model with a 1M context and a clear architecture gives students many testable questions:

• Long context: hide facts at different depths in a very long document and measure how often the model finds them. Does recall hold up near 1M tokens?
• Effort vs cost: run the same 50 problems at low, high and max reasoning effort. Plot accuracy against the tokens used.
• Indian languages: count tokens for the same sentences in English and several Indian languages with the 160K vocabulary. Compare with Sarvam 105B.
• Tools vs no tools: ask factual questions with and without web search, and check the answers by hand.
• Hybrid attention: Moonshot released smaller Kimi Linear models built on the same KDA layers. Study the design in a model you can actually run.

Write down the model version, settings and prompts for every test, so others can repeat it. To find teammates for projects like these, join APJ-RAISE. It is free.