As of November 4, 2026, Moonshot AI’s official repository offers public downloads of Kimi K3.1 Pro’s weights—the files holding the model’s learned parameters. For businesses and research teams, that provides a route to running and adapting the model on infrastructure they control.
The documented starting configurations use data-center hardware. Before commissioning a deployment, buyers need to establish which license terms apply and whether control over the model justifies operating it.
The initial repository commit is dated November 2. That does not establish when the files became public; the exact public-release time remains unverified.
The active parameter count is not the download size
Moonshot’s model card specifies 2.8 trillion total parameters, with 104 billion active for each token—a unit of text processed by the model. Its mixture-of-experts design selects only part of the model for each token. The smaller active count describes the computation for that step, not all the weights a deployment must accommodate.
The repository metadata lists 96 weight files and about 1.56 TB of stored data. That storage total is not a GPU-memory requirement: a running service also needs memory for handling requests and serving overhead.
Ask for a configuration, not just a GPU count
vLLM, software for serving models, documents a baseline with eight NVIDIA GPUs: two GB300 trays or one B300 node. The project says that baseline’s throughput was not tuned. This is a working example, not a universal minimum.
In its development-version Dynamo documentation, NVIDIA lists an eight-GB300 setup and a 16-B200 setup. These are hardware-specific deployments. A useful quote needs the accelerator model, memory, interconnect and expected workload—not simply a promise of eight GPUs.
Input support also depends on the chosen software configuration. Moonshot lists text and image inputs for the model, but NVIDIA labels the SGLang configurations on its Dynamo page experimental and text-only; image inputs use separate vLLM configurations. A business planning to process scanned documents should have image support specified in the proposal.
Likewise, Moonshot’s roughly one-million-token context window describes the amount of material the model can work with at once. It is not a promise of that capacity for every simultaneous user. Ask for document length, concurrent requests and response time to be evaluated together.
Which license conditions apply?
The custom license permits no-charge use and modification, with conditions:
Internal use. Sections 2–3 do not apply if third parties receive no access to the software, outputs or capabilities. Those sections also exempt use through Moonshot’s official products or certified inference partners.
Model-service businesses. If you or an affiliate operates one and combined group revenue exceeds US$20 million over any consecutive 12 months, a separate Moonshot agreement is required before commercial use, subject to those exemptions.
Large commercial products. Using the model in a product exceeding 100 million monthly active users or US$20 million monthly revenue requires displaying Kimi K3.1-Pro prominently in its interface, subject to the exemptions.
The model-service definition covers meaningful third-party control over inference or fine-tuning inputs, parameters or training data. It excludes capabilities solely embedded in specific end-user features and simple forwarding to another host. Notice-retention and other conditions still apply.
Check who receives outputs and what affiliates do against the full terms before relying on an exception.
Compare the service you actually need
For teams that simply want to use Kimi, Moonshot lists hosted access through its apps, Work, Code and API. Downloading weights does not supply a finished Kimi Work replacement: the company’s application examples use external tools and surrounding workflows. Include the interface and integrations in a self-hosting proposal.
The API pricing documentation describes usage-based billing for input, output and cache writes, separate from Kimi membership and Kimi Code. Exact token rates remain unverified, and the available evidence does not establish that self-hosting is cheaper.
For a meaningful cost comparison, request quotes for the same document lengths, simultaneous users and response-time target. Include storage, networking, power or rental, staffing, monitoring and idle capacity on the self-hosted side.
Buyers comparing other downloadable models can also read our Mistral Large 4 self-hosting guide. Its hardware guidance and license terms are separate from Kimi’s.
Our assessment: start with hosted evaluation when the immediate question is whether Kimi suits your work. Commission a self-hosting assessment when you can name a requirement for deployment control or adaptation that the hosted route does not meet. The next commitment should be a workload-specific evaluation and operating quote, not a hardware order based on the download alone.