Moonshot AI released the weights of its flagship model Kimi K3 on Monday. They came with a license of its own, no longer based on the previous “Modified MIT” pattern – and one that targets a single group above all: providers who resell the model as a service.

Anyone downloading Kimi K3 to run it internally, embed it in a product or use it for research can do so free of charge and without asking. But anyone running a “Model as a Service” business with it – making inference or fine-tuning available to third parties, via an API for instance – and generating more than 20 million US dollars in revenue together with affiliated companies over any twelve consecutive months must, according to the license text, enter into a separate agreement with Moonshot AI before any commercial use.

The sharpest clause in the new license is therefore not aimed at developers or corporate users, but at the layer above them: cloud and inference providers that host and resell open models. In other words, precisely those companies for whom the open-weight release of a frontier model is commercially most attractive.

What the “Kimi K3 License” says

Formally, the document is built as an MIT license: use, copying, modification, distribution, sublicensing and sale are explicitly permitted, as are deployment, fine-tuning and the creation of derivative works. Unlike classic software licenses, Moonshot explicitly names the subject matter as an AI artefact: model weights, parameters, configuration files as well as inference and training code all fall under the term “Software”.

Three conditions are layered on top:

Model as a Service (§2): If the licensee operates a MaaS business and exceeds the 20-million-dollar threshold, a separate agreement with Moonshot is required. By the license’s definition, MaaS applies when third parties can exercise “meaningful control” over inputs, parameters or training data. Exempted are end-user products in which model capabilities are embedded solely within specific features, as well as the mere relaying of requests to models hosted by others.

Branding (§3): Products or services with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must display “Kimi K3” prominently in the user interface.

Exemptions (§4): Both obligations fall away for purely internal use – defined as use that makes neither the software nor its outputs nor its underlying capabilities available to third parties – as well as for use via Moonshot’s official products or via “certified inference partners”.

The last exemption is strategically notable: it creates an incentive to obtain Kimi K3 through channels certified by Moonshot rather than self-hosting. Which providers qualify as certified is not defined in the license text. That also makes Moonshot AI’s aim clear: for smaller companies, and for internal use even at larger ones, Kimi K3 can be used free of charge. But anyone who wants to resell it, or build it into consumer products with significant reach (100 million-plus users), will most likely be asked to pay.

That effectively turns the open-weight model into a freemium product: free to get started, chargeable at heavier use. For the vast majority of users – internal use, research, startups, products with an embedded model – the license therefore behaves in practice like MIT. It becomes relevant for those selling Kimi K3 as a service.

Kimi K3 almost on a par with the US top models

Kimi K3 was unveiled on 16 July and is a mixture-of-experts model with roughly 2.8 trillion parameters, of which only 16 out of 896 experts are active per token – effectively around 50 billion. The context window spans one million tokens, complemented by native image processing and an always-on reasoning mode. Architecturally, Moonshot relies on two internally developed techniques, Kimi Delta Attention and Attention Residuals, which the company had previously published as open research. Following the launch, Moonshot had to temporarily pause new subscriptions because of GPU shortages.

In terms of capability, K3 performs at a level previously reserved for proprietary top-tier models. Across the common comparison indices it ranks in the leading group, narrowly behind Anthropic’s Claude Fable 5 and ahead of older frontier models. Its strengths lie clearly in coding and agentic tasks, while proprietary models remain ahead on broader reasoning and image understanding. On API pricing, K3 sits well below the US top-tier models, but above the cheaper open competitors from China.

The significance of the release lies less in individual benchmark points than in the combination of capability and availability: for the first time, a model of this class is available for download. The launch triggered immediate market reactions on the Nasdaq and calls into question the assumption that Chinese labs remain well behind the US frontier. And the more open weights at frontier level become a genuine alternative for hosters and enterprises, the more relevant the question of the conditions under which they may be used. That is precisely where the new license comes in.

Aus Datenschutz-Gründen ist dieser Inhalt ausgeblendet. Die Einbettung von externen Inhalten kann in den Datenschutz-Einstellungen aktiviert werden: