The weights of Kimi K3.1 Pro were officially released on November 4. Moonshot AI published the moonshotai/Kimi-K3.1-Pro repository on Hugging Face that day, keeping the promise made at its October 23 launch to release the weights by November 4. This is a genuine open-weights release: 2.8T total parameters, 104B activated parameters and a one-million-token context window, with a full evaluation table on the model card. But downloadable does not mean runnable for everyone — the deployment bar and the license terms are the parts that actually need explaining.
What is actually in the repository
According to the Hugging Face model card, Kimi K3.1 Pro is a natively multimodal agentic model using a mixture-of-experts architecture: 896 experts with 16 activated per token, 93 layers, a context length of 1,048,576 tokens, a 401M-parameter MoonViT-V2 vision encoder, MXFP4 weights and MXFP8 activations from quantization-aware training. The repository holds about 1.56TB across 96 shards, and the official recommendation is to run it with vLLM, SGLang or TokenSpeed. On evaluation, the model card reports 65.7% on Terminal-Bench 4.0, slightly below Claude Opus 5.5 at 66.4% and above GPT-6 Astra at 58.2%. Those are Moonshot AI's own numbers; independent reproduction on the open weights has not yet appeared, so the basis should be stated when quoting them.
Read the license and the deployment bar first
This release uses a dedicated Kimi K3.1 Pro License, not a standard MIT license. Under its terms, a model-as-a-service business with more than $20 million in revenue over 12 months needs a separate agreement with Moonshot AI, and products with more than 100 million monthly users or more than $20 million in monthly revenue must also name the model; the original text should be read before commercial use. On deployment, a 1.56TB download and a 2.8T-parameter model make this a server-class model that needs a multi-GPU environment and the operations capacity to run it — ordinary personal computers and single-GPU workstations are out of scope. For most users, staying with the Kimi apps and API remains the more realistic option, priced at $2 per million input tokens and $10 per million output tokens; the open weights mainly change the reproduction and customization space for research teams and teams with compute.
What the open weights change
For researchers and teams with infrastructure, open weights mean they can reproduce evaluations, fine-tune for specific needs and put the model into their own inference pipelines instead of accepting only the hosted service's version and pricing. For the ecosystem, open weights at the 2.8T scale are uncommon, and this release will become a reference point for testing large-model inference frameworks and quantization approaches. It does not change what Kimi K3.1 Pro is: first of all the flagship model of Moonshot AI's hosted product line, with open weights as another way of delivering the same product. What is worth watching next is the scores independent evaluators obtain on the open weights, and whether the community can actually run it at a reasonable cost.
