Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, meeting the 4 November date it promised at launch. The release puts one of the largest open-weight models in public hands, with 2.8 trillion parameters in total and 104 billion active for each token. The files are listed under moonshotai/Kimi-K3.1-Pro, with no access gate, and they run to about 1.56 TB across 96 files. Kimi K3.1 Pro is a post-trained version of Kimi K3, built with substantially larger post-training compute, according to Moonshot.
The model was launched on 23 October as the first in Moonshot's new K3.1 series and its most capable 3T-class open-weight model to date. It keeps a context window of one million tokens, or 1,048,576 tokens, and it takes text and images as input. The design is a mixture of experts with 896 experts, of which 16 are selected for each token, spread across 93 layers. The weights are stored at 4-bit precision, in MXFP4 form, after quantisation-aware training, with MXFP8 activations. A 401-million-parameter vision encoder, MoonViT-V2, handles the image side.
Kimi K3.1 Pro scores near Claude Opus 5.5
Moonshot's own benchmark table puts Kimi K3.1 Pro at 65.7 percent on Terminal-Bench 4.0, a test of coding agents on real terminal tasks. That sits just behind Claude Opus 5.5 at 66.4 percent and ahead of GPT-6 Astra at 58.2 percent. The numbers are Moonshot's own, published on the model card, and no third-party evaluation had been published by the time the weights went out. The model only works in thinking mode: requests with thinking turned off are rejected, according to the model's documentation. A faster version, Kimi K3.1 Pro Highspeed, runs at two to three times the speed of the first version, and it has no weight release of its own.
Running the model at home is not a realistic option for most people. The download is 1.56 TB, and the simplest vLLM setup needs eight data-centre graphics cards, according to the recipes published for the release. Moonshot recommends three inference engines, vLLM, SGLang and TokenSpeed, and leaves the hardware to their recipes. Most users will keep using the hosted version instead, through the company's app and coding tools or its API. The hosted model is listed on the company's platform under its usual name.
Kimi licence sets limits for large hosts
The licence is not a standard open-source licence. It is the Kimi K3.1 Pro Licence, a custom licence that follows the pattern of the K3 licence, word for word apart from the model's name. For most people it works like a permissive licence: they may use, copy, change, fine-tune and sell the model, as long as they keep its copyright notice. Two conditions bite at scale. A company running a model-as-a-service business with more than 20 million dollars in revenue over 12 months must sign a separate agreement with Moonshot before using it commercially. Products with more than 100 million monthly users, or more than 20 million dollars in monthly revenue, must show the model's name.
The release also sets a price for the hosted model. Moonshot lists Kimi K3.1 Pro at 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input at 20 cents per million. The Highspeed version costs 4 dollars in and 20 dollars out per million tokens. The earlier K3 cost 3 dollars in and 15 dollars out, so the new model is cheaper to run through the company's API. The price is 33 percent lower than before, according to The Decoder.
Weights arrive after July K3 launch
K3, the model this release builds on, launched on 16 July 2026 as a 2.8-trillion-parameter open-weight model with native multimodal input. The new version keeps the same size and the same million-token window, and adds the post-training that Moonshot says lifted its coding scores. The weights were promised by 4 November at the October launch, and Moonshot met that date. The company apologised on X on 25 October for rate limits and unnecessary quota use on the subscription plan, and gave every subscriber a one-time reset restoring 25 percent of their monthly quota.