Moonshot AI released Kimi K3.1 Pro as open weights under a short, permissive license. The legal terms are easy to live with. The 2.8-trillion-parameter download is the part that sets the limits.
The license for Kimi K3.1 Pro runs 3.08 kilobytes. The model it covers has 2.8 trillion parameters. That mismatch sums up the release: the legal permissions are short and generous, while the practical ones are decided by hardware.
A License Shorter Than Most Terms of Service
Moonshot AI put Kimi K3.1 Pro on its Hugging Face organization page this week. The model card describes an open-weight, natively multimodal agentic model, post-trained from Kimi K3 with a 1-million-token context window. Its license reads like an MIT-style grant. It covers the weights, parameters, configuration files, inference and training code and documentation, and it permits use, copying, modification, distribution, sublicensing and sale. Fine-tuning and derivative works are explicitly included.
The base conditions are light. The copyright notice has to travel with copies, and use must comply with applicable law. The software comes as is, with no warranty. For most developers, researchers and companies, that means they can download the weights, adapt them and ship products on top of them without negotiating anything.
The Two Revenue Triggers Hidden in the Conditions
The exceptions sit in the middle of the file and deserve a slow read. The first targets “Model as a Service” businesses, meaning those that give a third party access to inference or fine-tuning, through an API for example, in a way that lets that party meaningfully control inputs, parameters or training data. If a licensee and its affiliates run such a business and their combined revenue passes $20 million in any 12-month stretch, the license requires a separate agreement with Moonshot before commercial use.
The definition has carve-outs. End-user products with model capabilities embedded in specific features or harnesses don’t count, and neither does simply relaying requests to models hosted elsewhere. The second trigger concerns attribution. A commercial product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display “Kimi K3.1-Pro” in its interface.
Both triggers are waived for internal use that never exposes the model or its outputs to third parties, and for use through Moonshot’s official products or certified inference partners. The practical reading is straightforward: startups and internal teams are largely unconstrained, while big inference resellers and consumer-scale products carry obligations. Anyone close to those thresholds should have a lawyer read the text, since this is a summary and not legal interpretation.
What Open Weights Let You Do That the API Does Not
The weights open doors a hosted endpoint can’t. Teams can fine-tune the model on private data, run it inside their own network and keep prompts and outputs off third-party servers. Quantizing it is also possible, and the model page already lists 6 quantizations and 11 finetunes of K3.1 Pro. The card names vLLM, SGLang and TokenSpeed as recommended inference engines, each with its own setup recipe.
Control over serving makes a real difference. The card says its benchmark results use maximum reasoning effort, a temperature of 1.0 and top-p of 0.95, often through agent harnesses such as Claude Code or Kimi Code. On your own hardware, those settings are yours to choose. On the API, Moonshot serves the model its own way.
One caution: the card’s Transformers examples load the model with trust_remote_code=True, which runs code pulled from the repository. Reading that code first is wise, and serving through one of the recommended engines avoids the issue altogether.
The Real Download Is Measured in Terabytes
Anyone expecting a small download will be disappointed by the model card. K3.1 Pro is a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion active per token, spread across 896 experts with 16 selected for each token. Its weights use MXFP4 quantization with MXFP8 activations, trained with quantization-aware training.
Even at four bits per weight, the math puts the weights alone near 1.4 TB, before counting the vision encoder and other components. That figure is an estimate from the parameter count, not a number Moonshot publishes. An eight-GPU node with 80 GB per card offers 640 GB of memory, less than half of what’s needed. Active parameters set the compute per token, but every expert has to sit in memory somewhere, so running the full model takes a multi-node setup or very large memory pools.
Smaller community quantizations may shrink the footprint, but they change the model, and the benchmark numbers in the card don’t carry over to them.
Where the Hosted API Still Has the Advantage
The API has strengths of its own. Moonshot’s quickstart offers kimi-k3.1-pro through OpenAI- and Anthropic-compatible endpoints, with a reasoning_effort setting of low, high or max, defaulting to max. It supports vision input, structured output, tool choice and context caching.
The API also enforces a protocol detail. For multi-turn conversations and tool calls, the model expects the full assistant message, including reasoning_content and tool_calls, to be passed back unchanged. Self-hosters will need to reproduce that handling themselves.
Hosted access also sidesteps the license triggers. Use through Moonshot’s official products or certified partners is exempt from the commercial conditions, so a large company that would owe an agreement for self-hosted resale can often route through the API instead.
How to Decide Between Running It and Renting It
The license is the easy part. It is permissive, short and aimed mostly at a handful of very large operators. Infrastructure is the hard part. A small team gets the model at its best settings from the API with no hardware burden. A company that needs data to stay in-house, or wants to fine-tune, gets real value from the weights, but pays for it with a serious inference cluster.
Companies reselling inference at scale are the ones to watch. Whether Moonshot signs separate agreements with them, and on what terms, will show how open the release is in practice. For everyone else, the headline holds: the weights are free to use, but the model’s size decides who can actually run them.
Disclosure: This is not trading or investment advice. Always do your research before buying any cryptocurrency or investing in any services.
Follow us on Twitter @themerklehash to stay updated with the latest Crypto, NFT, AI, Cybersecurity, and Metaverse news!