Moonshot AI releases Kimi K3 model weights under custom licence: What it means for enterprises

Beijing-based Moonshot AI completed the rollout of Kimi K3, its largest and most performant open-weight AI model, by releasing the model weights under its own licence.

The Chinese AI startup on Monday, July 27, also published a 47-page technical report comprising details about the techniques used to train Kimi K3, what obstacles it encountered, and the infrastructure required to run the open-weight AI model independently.

Kimi K3 is a 2.8 trillion parameter model with a one million-token context window and Mixture-of-Experts (MoE) architecture that was, so far, accessible via Moonshot’s hosted API since its debut earlier this month. The startup promised at the time to make the model’s weights,  billions of individual settings acquired and fine-tuned during training that help determine how a model responds to prompts, freely available for everyone under an open-source licence.

While Moonshot AI followed through on its commitment, the weights of the model have been released under a custom ‘Kimi K3’ licence instead of traditional open-source licences such as MIT and Apache 2.0.

Keeping the weights of a large language model (LLM) secret, like OpenAI and Anthropic, restricts developers and researchers from taking a look inside or deploying the model on their own infrastructure. In contrast, open-weight models allow developers to download, modify, and run on their own servers or devices subject to certain conditions laid down as part of the licensing terms.

Given that the technical documentation published by Moonshot AI also includes details about the inference infrastructure, optimised attention kernels, MoE communication libraries and deployment components, it could enable researchers and enterprise developers to self-host the model more easily rather than consume it exclusively through an API.

Kimi K3 has been turning heads in Silicon Valley since its launch as it is said to achieve performance on par with leading US models like Anthropic’s Claude Fable 5 while coming out to be cheaper to use.

Kimi K3 API access, for instance, is priced at $3 per million input tokens and $15 per million output tokens. Beyond Kimi K3, recent data suggests that 30-46 per cent of enterprise token usage on Chinese open-weight AI models is by US businesses via LLM gateway platforms such as OpenRouter.

Note, Anthropic and US government officials have accused Moonshot AI of using US-developed models like Fable 5 to train Kimi K3 via distillation, a widely used ⁠AI ​training technique in which developers use the outputs of a more capable model to train or improve another model. US officials maintain there is a distinction between legitimate distillation and large-scale extraction that amounts to intellectual property theft.

What the Kimi K3 licence allows

Experts have pointed out that the Kimi K3 licence is largely inspired by the permissive MIT licence. However, it also imposes some obligations on larger companies and AI service providers that are not found in traditional open-source licences.

For instance, if the licensee or any of its affiliates is a ‘model as a service’ business whose annual aggregate revenue exceeds $20 million USDthen the licensee must enter into a separate agreement with Moonshot AI before using the software or its derivative works for any commercial purpose.

The company defines ‘model as a service’ as any business that provides “third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data.” “This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others,” as per the licence.

Another obligation under the Kimi K3 custom licence involves branding. If a licensee with more than 100 million monthly active user or generating more than $20 million USD in monthly revenue, uses any open version of Kimi K3 for commercial products or services, then the licensee must prominently display ‘Kimi K3’ as a label on the user interface of such products or services.

This requirement essentially ensures that companies which typically abstract away the underlying model will need to disclose it directly within their products, as per the licensing terms.

There are also certain exceptions. Companies looking to use Kimi K3 only for internal use, where the models’ outputs are not made available to third parties, are exempted from the above-mentioned obligations. Additionally, Kimi K3 usage via Moonshot AI’s official products or certified inference partners is exempted.

Moonshot AI is not the first AI company to adopt such a licensing strategy. Meta’s Llama family of open-weight models have long been distributed under its own community licence which similarly requires a commercial agreement for those building with any of the models and exceeding 700 million monthly users.

Comments are closed.