Kimi K3 license complicates Moonshot AI’s open-weights release

Moonshot AI has released the full weights for Kimi K3, its largest Kimi model, along with a Kimi K3 license that gives developers broad access but adds conditions for some commercial users. The Chinese AI startup is offering the 2.8 trillion-parameter Mixture-of-Experts model for self-hosting, a move that matters for enterprises looking for frontier-class models they can run outside a hosted API.

The release completes a rollout that began with Kimi K3 on Moonshot’s hosted API earlier this month. Moonshot also published a 47-page technical report and deployment components, including inference infrastructure, optimized attention kernels, MoE communication libraries and support for vLLM and SGLang. The company says Kimi K3 activates 104 billion parameters from 896 experts, supports a one million-token context window and includes native multimodal reasoning.

What does the Kimi K3 license allow?

The license allows users to download, use, modify, distribute, deploy, fine-tune and sell copies or derivative works of the software, including model weights, parameters, configuration files, code and documentation. That makes Kimi K3 accessible to researchers, startups and many enterprises, but it is not the same as a standard Apache 2.0 or MIT-style open-source license.

The main commercial caveat applies to companies that operate what Moonshot defines as a “Model as a Service” business. Under the license, if a licensee and its affiliates exceed $20 million in aggregate revenue over any consecutive 12 months, they must sign a separate agreement with Moonshot before using Kimi K3 or derivatives for commercial purposes in that kind of service.

Moonshot defines Model as a Service as giving third parties access to model inference or fine-tuning, such as through an API, where those third parties have meaningful control over inputs, parameters or training data. The definition excludes end-user products where model capabilities are embedded in specific features, and excludes relaying requests to models hosted by others.

The revenue language is broader than product-specific revenue. It refers to revenue of the licensee and affiliates, which means a smaller business unit could fall within the requirement if its parent company crosses the threshold and uses Kimi K3 in a qualifying service.

Which enterprise uses trigger extra terms?

Internal use has a carveout. Moonshot’s license says the commercial licensing and attribution provisions do not apply when the model is used only inside an organization and does not make the software, outputs or underlying capabilities available to third parties.

That distinction could matter for banks, retailers or other non-AI companies using Kimi K3 for employee productivity, internal search or customer-service features. Companies selling model access, fine-tuning tools or API-based AI services have more to review, especially if they or their affiliates clear the $20 million annual revenue threshold.

The license also adds a branding requirement for large commercial products. If Kimi K3 or a derivative is used in a commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, “Kimi K3” must be displayed prominently in the user interface.

Developers focused on the open-source caveat

AI researcher Nathan Lambert, previously co-leader of Ai2’s Olmo open-model work, wrote on X that the license is MIT-inspired but “distinctly non-commercial” for companies above $20 million in annual revenue, while also requiring Kimi K3 display at larger scale.

Developer reaction also focused on the scope of the release. Moonshot published not only weights, but supporting infrastructure such as attention kernels, MoE communication libraries and agent tooling. At the same time, discussion has centered on whether Kimi K3 is better described as open weight rather than fully open source because of the commercial conditions.

Operating the model is another constraint. The weights are roughly 1.5 TB, putting Kimi K3 in reach mainly for organizations with substantial inference infrastructure, even as reports surfaced of deployments using clusters of consumer RTX 5090 GPUs.

Moonshot is following a wider pattern among frontier AI developers. Meta’s Llama models, for example, use a community license that requires a commercial agreement for users above 700 million monthly users rather than a traditional open-source license. For enterprises, Kimi K3 adds one more reminder that model benchmarks and model licensing now need to be evaluated together.

This story draws on original reporting from VentureBeat.