Kimi K3.1 Pro Weights Land at 2.8T Parameters, With a Licence That Bills at $20 Million

Moonshot AI put the weights of Kimi K3.1 Pro on Hugging Face on Wednesday, nearly two weeks after the model went live on its paid API. The model card lists 2.8 trillion parameters in total and 104 billion active per token, under a custom licence rather than MIT or Apache 2.0.

Moonshot released the model through its API on October 23, 2026, and promised weights for November 4, according to LetsDataScience. The card does not give a release date. It does not say whether the download is gated, and it does not list hardware requirements. It recommends vLLM, SGLang or TokenSpeed for serving.

What the model card lists

K3.1 Pro is a mixture-of-experts model with 93 layers, one of them dense. It has 896 experts, selects 16 per token and keeps 2 shared. The context window is 1,048,576 tokens. A 401-million-parameter MoonViT-V2 encoder handles images. Weights use MXFP4 quantisation with MXFP8 activations.

Moonshot says it post-trained the model from Kimi K3 with large-scale reinforcement learning. Against K3, it swapped the full-attention Gated MLA layers for Gated DeepSeek Sparse Attention layers. The card says nothing about training data.

The licence has two revenue triggers

The Kimi K3.1 Pro License permits commercial use, with conditions that depend on size. The summary below is from the licence text on the Hub. Anyone shipping a product should read the full file.

Trigger	Threshold	Obligation
Model-as-a-service business	Revenue above $20 million over 12 months	Separate agreement with Moonshot AI
Commercial product	More than 100 million monthly users, or more than $20 million monthly revenue	Display "Kimi K3.1-Pro" in the interface
Internal use, or access through Moonshot products and certified inference partners	None	Exempt from both

Moonshot's licence defines model-as-a-service as giving a third party inference or fine-tuning access with meaningful control over inputs, parameters or training data. The licence points questions to license@moonshot.ai.

Moonshot's numbers and the independent ones

Moonshot's benchmark table is self-reported. It uses reasoning effort set to max, temperature 1.0 and top-p 0.95. Its headline scores include 94.2 on GPQA-Diamond, 92.1 on BrowseComp and 65.7 on Terminal-Bench 4.0, the last run through Claude Code.

Vals AI ran its own suite with the same sampling settings. It scores K3.1 Pro at 62.52% on the Vals Index, plus or minus 1.14, ranking it first among open-weight models. DeepSeek V4.1 Flash 1017 scores 61.26%. On Terminal-Bench 4.0 Vals measured 63.64%, third overall and about two points behind Claude Opus 5.5.

1. Kimi K3.1 Pro62.52 %
2. DeepSeek V4.1 Flash 101761.26 %

Source: Vals AI, accessed 2026-11-04

The gap between the two open models is 1.26 points, close to the 1.14-point margin on Vals' score for Kimi. Vals also lists weak spots. K3.1 Pro placed 43rd of 89 on Legal Research Bench at 37.50% and 42nd of 80 on Tax Agent Bench at 24.87%, and Vals says it trails Kimi K3 on those and on HLAB-Vals-1.1. Vals lists the API at $2 per million input tokens and $10 per million output tokens, and records $6.44 per test run on the index.

Vals' page gives two rank counts for the index, 9 of 32 and 8 of 27, which suggests it was updated after first publication. The later figure is the one that places K3.1 Pro first among open weights.

What is not known

Moonshot has not published training data details, hardware sizing or the weight file size on the card. No independent party has reported running the model locally. Vals tested it through the API, so its scores say nothing about how the quantised download behaves on a self-hosted cluster.

Moonshot has not said when a K3.1 variant smaller than 104 billion active parameters will follow. Earlier open releases on the site, such as Baidu's Unlimited-OCR-2 and Nvidia's Nemotron 3.5 Transcribe, shipped with simpler terms and far smaller footprints.