Kimi K3.1 Pro weights arrive, bringing a 1.56TB deployment and a custom commercial license

Moonshot AI announced the public weights for Kimi K3.1 Pro on November 4, following the model’s October hosted launch. GIGAZINE’s November 5 coverage records the release announcement, and Moonshot’s Hugging Face repository now contains the downloadable files. Developers can evaluate a self-hosted deployment rather than rely solely on API access. Dated independent coverage, Official model repository
The practical question is who can run it economically. The release expands access to the model’s underlying weights, but its scale and license require a more deliberate assessment than downloading a small desktop assistant.
Sparse computation, substantial storage
Moonshot’s model card lists 2.8 trillion total parameters and 104 billion activated parameters, with text and image input and a 1,048,576-token context limit. The model uses a mixture-of-experts architecture and quantization-aware training, with MXFP4 weights and MXFP8 activations. Model specifications
The official file listing shows approximately 1.56TB across the repository, including weights split into 96 safetensors shards. That is the distribution’s size, not a measurement of the memory needed to serve requests. Download contents
Active parameters describe how much of the network participates in computation. They do not establish a complete hardware budget. A deployment still has to accommodate its chosen weight representation, working memory, request history and concurrent users. Treating the active count as though it were the entire model would obscure the main operational constraint.
AMD’s November 4 technical article reports serving the released checkpoint on eight Instinct MI355X GPUs through ATOM and vLLM. Its measurements use a particular software stack and explicitly are not performance guarantees. This is a documented server configuration, rather than evidence that the model is practical on an ordinary laptop. AMD deployment report
Commercial permission has conditions
The custom Kimi K3.1-Pro License permits use, modification and redistribution subject to conditions. A licensee operating a Model as a Service business, including through affiliates, needs a separate agreement if aggregate licensee-and-affiliate revenue exceeds $20 million over any consecutive 12 months. The requirement applies before commercial use of the software or derivatives. License text
A separate clause requires prominent model branding on commercial products exceeding 100 million monthly active users or $20 million in monthly revenue. The license specifies exceptions for its defined internal use and access through Moonshot’s official products or certified inference partners. Organizations should check the definitions against their actual offering. Commercial conditions and exceptions
Evaluate the deployed system
Moonshot recommends vLLM, SGLang and TokenSpeed. Its benchmark tables generally use maximum reasoning effort, with specified sampling settings. Those conditions matter when comparing published scores with a lower-latency production configuration. Serving guidance and evaluation conditions
A useful trial should test representative documents, coding tasks and tool calls while tracking completed work, latency and total operating cost. Long context should be evaluated for accurate retrieval and sustained reasoning, not simply whether a large input is accepted.
The weight release makes direct inspection and controlled deployment possible. Whether that control is worth the infrastructure and licensing obligations depends on the workload, the serving setup and the alternatives already available to the team.
Sources and citations
Revision history
Published .