Moonshot AI published the full weight set for Kimi K3 on July 27, eleven days after the Beijing lab launched the model as a hosted service. The release runs to 2.8 trillion total parameters, of which 104 billion activate on any given token, and pulling the files from Hugging Face means downloading 1.56 terabytes. Moonshot calls the Kimi K3 license “open weight,” not open source, and that distinction is not cosmetic. It changes what a reseller owes once the model starts making money.
What Is the Kimi K3 License?
The Kimi K3 License is Moonshot AI’s own custom terms for the model weights, not a standard open source license. It grants free use, modification, redistribution and derivative works, then attaches two separate conditions. Where a licensee or any of its affiliates operates a Model-as-a-Service business and that group’s aggregate revenue exceeds $20 million over any consecutive twelve months, the licensee must enter a separate agreement with Moonshot AI before using the software. Separately, products with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display “Kimi K3” prominently in the interface.
Specs Without the Marketing Layer
Strip away the release-day framing and the numbers are still large. Kimi K3 is a mixture-of-experts model that routes each token to 16 of 896 experts, which is how it reaches 104 billion active parameters out of 2.8 trillion total without running every parameter on every request. The context window holds 1 million tokens. Input is text and images, processed through a vision encoder Moonshot calls MoonViT-V2; there is no confirmed video support in the current release.
What Moonshot did not publish matters as much as what it did. Training data and training code stayed private. Only the weights and the technical report went out, so anyone reproducing the model’s behavior, or auditing what it learned from, is working from the outputs alone. It is worth stating plainly rather than assuming “open” covers the whole pipeline.
Two Thresholds, Two Different Obligations
The license does not gate access. It gates growth, and it does so through two separate triggers that kick in at different points and demand different responses.
| Threshold | Obligation |
|---|---|
| Licensee or affiliate operates a MaaS business and that group’s aggregate revenue exceeds $20M over any consecutive 12 months | Enter a separate agreement with Moonshot AI before using the software |
| More than 100M monthly active users, or more than $20M in monthly revenue | Display “Kimi K3” prominently in the product interface |
Simon Willison, writing on his blog the day the weights dropped, noted that Moonshot makes no attempt to describe this as an “open source” license, using “open weight” throughout instead. Coverage from unite.ai walks through the license text and describes it as reading like MIT for most of its length, right up until the revenue clauses appear.
Why the Hosting List Matters More Than the Download Link
Six vendors announced inference availability the same day the weights went public: Modal, Together AI, Nebius, GMI Cloud, Baseten and Fireworks AI. In practice, “open weight” here means renting Kimi K3 from one of those vendors rather than standing up 1.56 terabytes of infrastructure yourself.
What This Means for a Data Team
The question worth asking before adopting Kimi K3 is not whether it is free. It is at what scale your rights change, and whether your product’s growth trajectory puts you near either threshold within the next year rather than the current quarter. That framing echoes a pattern showing up elsewhere in the stack: Cloudflare and beehiiv’s AI Crawl Control turned AI access into a negotiated, measurable resource rather than a binary allow-or-block decision, and Moonshot’s revenue tiers do something similar for model licensing. Both replace a yes-or-no policy with a number you have to track.
For marketing and analytics teams already watching AI decisioning move into the core of the marketing stack, the licensing question is not academic. A vendor building on Kimi K3 inherits these thresholds, and a data team evaluating that vendor should ask where the vendor sits against them, not just what the model scores on a leaderboard.