Wednesday, 4 November 2026NewsletterAboutInstagramX
Subscribe

Moonshot Puts Kimi K3.1 Pro's 2.8T Weights Online

China's Moonshot AI has published the full weights of its top Kimi model, free to run and sell, with strings for very large users.

A row of steel-fronted computer racks with blue lights and power cables overhead in a data centre
File photo: server racks at Oak Ridge National Laboratory. Oak Ridge National Laboratory, CC BY 2.0, via Wikimedia Commons

[ The short version ]

  1. 01Moonshot AI has put the full weights of Kimi K3.1 Pro on Hugging Face, meeting the 4 November date it set when it launched the model in October.
  2. 02The model has 2.8 trillion parameters, of which 104 billion are used for each token, reads up to about 1 million tokens and handles images as well as text.
  3. 03Anyone may use, fine-tune and sell it, but AI-hosting firms with more than $20 million in yearly revenue need a separate deal, and apps with over 100 million monthly users must show the Kimi name.

Moonshot AI, the Beijing company behind the Kimi chatbot, has published the full weights of Kimi K3.1 Pro on Hugging Face, so developers can now download the model and run it on their own servers instead of only renting it through Moonshot's API. The company had promised the release by 4 November when it launched the model in October, and it announced the upload on X with the line "We'll keep advancing open intelligence", RuntimeWire reports.

When it launched, independent testers at Artificial Analysis ranked K3.1 Pro fourth on their Intelligence Index, behind two Claude models and Meta's Muse Wonder Full, which makes it the highest-scoring model whose weights anyone can download (TheStak covered the launch).

3T
total parameters in Kimi K3.1 Pro, with 104 billion active for each tokenMoonshot AI model card on Hugging Face

What is in the download

According to the model card, K3.1 Pro is a mixture-of-experts model: it has 2.8 trillion parameters spread across 896 expert networks, but switches on only 16 of them, or about 104 billion parameters, for each token it produces. That keeps it cheaper to run than its size suggests. It reads up to 1,048,576 tokens at once, and a built-in vision encoder lets it work with images as well as text.

The weights are stored in a compact 4-bit format that Moonshot says the model was trained to use, and the 96 files on Hugging Face still add up to about 1.56 terabytes. Running it means a cluster of data-centre GPUs, not a laptop. Moonshot points developers to the open-source serving tools vLLM and SGLang, which already have set-up guides for the model.

More from AI and Tech StakGoogle Ties School Search to AI, and the Backlash Is LoudK-12 schools that keep Gemini switched off would lose Google Search too. Senators, teachers and New York's mayor call it extortion; Google has softened it.

Free, with two conditions

The Kimi K3.1-Pro licence is close to the permissive MIT licence: anyone may use, copy, modify, fine-tune, distribute and sell the model and products built on it, free of charge. It adds two conditions aimed at the largest users.

  • A company that sells "model as a service", meaning API access that lets customers control the model's inputs or training, must sign a separate agreement with Moonshot before commercial use once its group revenue passes $20 million over any 12 months.
  • A commercial product with more than 100 million monthly users, or more than $20 million in monthly revenue, must show "Kimi K3.1-Pro" prominently in its interface.
  • Neither condition applies to internal use, where the model and its output are not offered to outsiders, or to use through Moonshot's own products and certified partners.

Moonshot has also said that it removed cyber-security training data from this released version, so that the downloadable model's ability to find software vulnerabilities stays behind leading closed models. That is the company's own account and has not been tested independently.

  • #MoonshotAI
  • #Kimi
  • #open-weightmodels
  • #AImodels
  • #ChinaAI

Written by , EditorView editor profile

Next on the Stak

Picked for readers of this story

Keep reading