Moonshot AI publishes Kimi K3.1 Pro weights, a 2.8 T total, 104 B active multimodal MoE
The weights ship in MXFP4 under Moonshot's own licence, which adds a revenue threshold for hosting providers. Context is 1 M tokens and the card recommends vLLM, SGLang or TokenSpeed.
Moonshot AI has put the weights of Kimi K3.1 Pro on Hugging Face, TechNode reported on 5 November 2026. The model card describes a mixture-of-experts transformer with 2.8 T total parameters and 104 B active per token, 93 layers, 896 experts with 16 routed and 2 shared per token, a 160 k vocabulary and a 401 M parameter MoonViT-V2 vision encoder for image input. Context length is 1,048,576 tokens. The weights ship in MXFP4 with MXFP8 activations, produced by quantisation-aware training rather than post-training conversion. The card gives no release date.
The licence is Moonshot’s own Kimi K3.1-Pro License. It is MIT-style with two added conditions: a model-as-a-service business whose revenue exceeds $20 million in any twelve months needs a separate agreement before commercial use, and products with more than 100 million monthly active users or $20 million in monthly revenue must display the model name in their interface. Internal use is exempt. This is open weights with restrictions, not open source.
The vendor reports 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, 92.1 on BrowseComp and 94.2 on GPQA-Diamond, with Claude Opus 5.5 and GPT-6 Astra listed as comparators on most rows. Recommended serving stacks are vLLM, SGLang and TokenSpeed, and TechNode notes that Huawei has announced support through vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 systems.
Why it matters
Few teams will host a 2.8 T parameter model, so the practical interest is in what the card reveals: an attention stack that mixes 69 KDA layers with 24 Gated DSA layers, and MXFP4 weights trained that way from the start, both of which are likely to appear in smaller models. The benchmark table compares against proprietary systems run under different harnesses, and the footnotes admit one internal result disagreed with a published one, so treat the numbers as indicative. Watch whether the licence’s revenue threshold deters hosting providers, since that decides how widely available the model becomes.