What is Kimi K3.1 Pro: 3 things on the 2.8T open model

Kimi K3.1 Pro is Moonshot AI's 2.8T open-weight model, with all 96 weight files on Hugging Face. Here are 3 facts to know before you download it.

Adam Wang

09/11/2026 · 7 min read

Facebook

What is Kimi K3.1 Pro: 3 things on the 2.8T open model

In this article (6)

Fact 1: 2.8 trillion parameters, genuinely open
Fact 2: an open license with a 20 million USD condition
Fact 3: a multi-GPU server model, not a laptop model
Benchmarks the vendor measured
Who should download, who should use the API
Sources

Kimi K3.1 Pro is Moonshot AI's newest open model, with 2.8 trillion total parameters and 104 billion active per token. All 96 weight files are live on Hugging Face, adding up to 1.56 TB. This article covers 3 facts to know before you download it: the real scale of the model. A license with a 20 million USD revenue condition, and why this is a multi-GPU server model rather than a laptop model.

I wrote this piece strictly from what the model's Hugging Face page and its license state, read on 08/11/2026. The question What is Kimi K3.1 Pro has a short answer. I haven't run the model myself, so you'll find no claims here about speed or answer quality from personal testing. Every figure carries a source and a read date at the end.

Fact 1: 2.8 trillion parameters, genuinely open

Kimi K3.1 Pro is a natively multimodal Mixture-of-Experts model that takes text and images. Its Hugging Face page calls it the company's most capable 3T-class model to date. Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. It's the first model in the K3.1 series.

The official spec sheet lists these numbers. Many readers ask What is Kimi K3.1 Pro and where to get it. Total parameters: 2.8 trillion. Active parameters per token: 104 billion. Layers: 93, with 1 dense layer. Experts: 896, with 16 picked per token plus 2 shared experts. Vocabulary: 160K tokens. Context length: 1,048,576 tokens, or about 1 million. The vision encoder is MoonViT-V2 at 401 million parameters. Weights use MXFP4, activations use MXFP8, and training was quantization-aware.

The weight repo on Hugging Face holds 96 safetensors files at 1.56 TB in total. The repo was created on 02/11/2026 and last updated on 05/11/2026, per Hugging Face API data read on 08/11/2026. The model page lists 22,571 downloads last month.

The model offers three reasoning effort levels: low, high, and max. Max is the default. All evaluation results on the model page run at max effort, temperature 1.0, and top-p 0.95. You can reach the model through Moonshot AI's platform by picking kimi-k3.1-pro, with OpenAI- and Anthropic-compatible APIs. For multi-turn chats and tool calls, the docs ask you to pass back the full assistant message. Including reasoning_content and tool_calls, not just the content field.

If you want another open model to compare it with, read What is Hermes Agent to see a different open-source agent approach. If you want a packaged model instead, read What is Gemini 3.

Chess pieces on a board, an illustration for the reasoning skill of Kimi K3.1 Pro — Chess pieces on a board, an illustration for the model's reasoning skill. Photo: Tima Miroshnichenko, Pexels, 2021-01-25.

Fact 2: an open license with a 20 million USD condition

Kimi K3.1 Pro ships under its own license, called the Kimi K3.1 Pro License, not Apache 2.0 or MIT. The license permits use, copying, modification, merging, publishing, distribution, sublicensing, and selling copies, including running, deploying, and fine-tuning the model. Two sections carry conditions that commercial users need to read first.

Section 2 covers Model as a Service, which means giving a third party access to model inference or fine-tuning through an API. If you or your affiliates run that kind of business, and your combined revenue tops 20 million USD across any consecutive 12 months. You must sign a separate agreement with Moonshot AI before using the software or its derivatives for any commercial purpose.

A desk with a laptop and papers, an illustration for reading the license carefully before commercial use. Photo: Kampus Production, Pexels, 2021-05-05.

Section 3 adds a branding threshold. If the software powers a commercial product or service with more than 100 million monthly active users. Or more than 20 million USD in monthly revenue, then the name "Kimi K3.1-Pro" must appear prominently on that product's user interface.

Neither section applies to internal use, which the license defines as use that doesn't make the model. Its outputs, or its capabilities available to third parties. They also don't apply when you access the model through Moonshot AI's official products or certified inference partners. In short: individuals who download it to test and research don't hit the 20 million USD bar. While companies that build an API service on top of it need a separate deal once revenue crosses it.

That sets it apart from Apache 2.0, the license many open models use, because Apache 2.0 puts no revenue bar on commercial use. If you compare open models by license, What is Claude Cowork shows how a commercial agent product frames multi-step work instead.

Fact 3: a multi-GPU server model, not a laptop model

This is the easiest point to get wrong, so I'll state it plainly: Kimi K3.1 Pro is a multi-GPU server model, not a laptop model. The full weights weigh 1.56 TB across 96 files. A normal laptop doesn't have the memory to load this model, even in MXFP4 format. When you read the word "open" here, it means the weights and code are public. While the hardware to run them is still a multi-GPU server.

The model page names three recommended inference engines: vLLM, SGLang, and TokenSpeed, with links to each engine's recipes. All three run on GPU clusters, not on end-user desktop apps. If you don't own a GPU cluster, you can still use the model through Moonshot AI's API by selecting kimi-k3.1-pro on platform.kimi.ai.

So the practical split is simple. Use the API if you want to try the model now without hardware. Download the weights if you run a GPU cluster and want to self-host. Fine-tune on your own data, or study a 2.8T MoE design up close. This article gives no specific GPU count, because the vendor publishes no minimum setup in the docs I read.

A laptop showing source code, an illustration for using Kimi K3.1 Pro to write code — A laptop showing source code, an illustration for using the model to write code. Photo: Rodrigo Santos, Pexels, 2020-03-08.

Benchmarks the vendor measured

The model page compares Kimi K3.1 Pro against five rivals: Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017, and GLM-5.3. Here are 8 representative scores, taken exactly from the vendor's table.

Benchmark	Kimi K3.1 Pro	Notes
Terminal-Bench 4.0	65.7	Claude Code harness
DeepSWE v1.1	75.5	Claude Code harness
SWE Marathon v1.1	51.7	Claude Code harness
BrowseComp	92.1	With tools, 500-step cap
GPQA-Diamond	94.2	Reasoning and knowledge
HLE-Full	51.1	Reasoning and knowledge
OSWorld V2.1	79.6	Kimi Agent harness
ExploitGym	20.7	Kimi Code, 6 hours per task

Typing code on a laptop keyboard in a dark room, an illustration for the model's coding benchmarks. Photo: Pavel Danilyuk, Pexels, 2020-09-19.

Two agent scores stand out: BrowseComp at 92.1 and HLE-Full with tools at 70.0. In office work, OfficeQA Pro v2 scores 64.4 and Automation Bench-AA scores 73.3, both the best in the vendor's table. In multimodal work, GDP.pdf with Python scores 85.7 and MMMU-Pro with Python scores 90.0.

Moonshot AI measured and published all of these scores itself, with no independent check yet. The coding tests run on the Claude Code harness, while ExploitGym runs on Kimi Code with a six-hour budget per task. Different harnesses give different scores, so use this table for rough comparison inside one table only. And don't treat any single number as a final verdict.

Who should download, who should use the API

Download all 96 weight files if you run a GPU cluster, want to self-host the model. Fine-tune it on your own data, or study a 2.8T MoE design. Read the license first, above all Section 2 on the 20 million USD revenue bar for Model as a Service.

A laptop showing source code, an illustration for weighing self-hosting against the API. Photo: Rodrigo Santos, Pexels, 2020-03-08.

Use the API if you want to test the model now without hardware. The API path on platform.kimi.ai lets you pick kimi-k3.1-pro and call it on OpenAI- or Anthropic-compatible endpoints. API use through the vendor's official products also stays clear of the separate-agreement condition in the license.

Don't download if you only have a laptop or a standard desktop. At 1.56 TB, the model needs a multi-GPU server, so a download in that case just eats disk space without running. If you enjoy seeing AI models make video, read What is Pika AI for a lighter topic after this spec-heavy one.

Learn with focus. Work with a goal. What is Kimi K3.1 Pro? This article answered that above. The clearest goal is to try the API first, read the license, and only then think about self-hosting.