Kimi K3.1 Pro open weights on BenchGecko · Kimi K3.1 Pro License on Hugging Face
Moonshot AI published open weights for Kimi K3.1 Pro on Hugging Face under the Kimi K3.1 Pro License. BenchGecko updates the row to open weights · 2.8T total and 104B active parameters · 1,048,576 token context.
What changed
Moonshot AI published the weights for Kimi K3.1 Pro on Hugging Face as moonshotai/Kimi-K3.1-Pro, meeting the November 4, 2026 date it set at launch. The API model has been listed on BenchGecko since 2026-10-23. BenchGecko updates the row from proprietary to open weights and sets the Hugging Face id.
vLLM published day 0 support for the release, covering NVIDIA GPUs, AMD GPUs, Google TPUs and Alibaba T-Head Zhenwu PPUs.
Specs
From the Hugging Face model card:
- Architecture · mixture of experts, 2.8T total parameters, 104B active per token
- Experts · 896, with 16 selected per token and 2 shared
- Layers · 93, mixing 69 KDA and 24 Gated DSA attention layers
- Context · 1,048,576 tokens
- Inputs · text and image, with a 401M parameter MoonViT V2 vision encoder
- Weights · MXFP4 weights with MXFP8 activations, trained quantization aware
- Licence · Kimi K3.1 Pro License
- Inference · vLLM, SGLang and TokenSpeed recipes
License terms
The Kimi K3.1 Pro License permits use, modification, redistribution and commercial use with notice. Two thresholds apply for large scale commercial use: a model as a service business above $20 million revenue over any consecutive 12 months needs a separate agreement with Moonshot AI, and a commercial product above 100 million monthly active users or $20 million in monthly revenue must show the name Kimi K3.1-Pro prominently. Internal use and access through Moonshot AI products or certified partners are exempt.
Pricing on BenchGecko
First party rates are unchanged: $2.00 input and $10.00 output per million tokens. BenchGecko keeps those price fields as they are and only flips the open weights flags.
Vendor reported benchmarks
Moonshot AI reports these results on the model card at reasoning effort max. BenchGecko records them as vendor claims.
- Terminal Bench 4.0 · 65.7
- SWE Marathon v1.1 · 51.7
- DeepSWE v1.1 · 75.5
- HLE Full with tools · 70.0
- BrowseComp · 92.1
- GPQA Diamond · 94.2
- OSWorld V2.1 · 79.6
- APEX Agents · 69.4
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.