Kimi K3: The License That Is Only Open Source in Name

There is no free lunch. Moonshot AI, a Chinese artificial intelligence lab, has just made its entire Kimi K3 model available for free download, touted as the largest ever released in open access. However, upon opening the accompanying file that typically comes with such releases, developers were met with a surprise. It's not the usual license that allows for nearly anything without compensation. Instead, it's a proprietary text that leaves small entities alone but sends the bill as soon as a company starts making significant profits.

Key Points

• Moonshot AI has released its entire Kimi K3 AI model for free, but under a proprietary license, not the standard open license.
• Any company reselling access to this model and exceeding $20 million in revenue over twelve months must sign a commercial agreement with Moonshot.
• After reaching 100 million active users per month, the name "Kimi K3" must be visibly displayed in the product using it.
• Strictly internal use within a company, without reselling to clients, escapes these two constraints.

Downloading an AI model means downloading its trained "brain".

To understand the matter, one must first grasp what Moonshot has actually put online. An AI model like Kimi K3, ChatGPT, or Claude relies on billions of internal adjustments, known as its "weights": this is the acquired memory of the model after months of training, enabling it to understand a question and respond. Downloading the weights of a model means directly retrieving its functional "brain", ready to run on its own machines rather than depending on an online service.

This is precisely what Moonshot has made public: the complete files of Kimi K3, over 1.5 terabytes of data (equivalent to several hundred high-definition films), available for download by anyone. The model also shows very good results on standard industry tests, comparable to the best paid models on the market like Claude or GPT. Technically, there’s nothing to criticize: it’s an achievement. The problem lies elsewhere, and it is purely contractual.

"Open source," really? Not quite.

In AI jargon, a model "open source" in the strict sense can be used, modified, and resold without any restrictions, by anyone, for any purpose. This promise has built the reputation of Chinese models in recent months, including Kimi K3 at its announcement.

However, the license published alongside the files tells a different story. The text introduces a very specific category: "Model as a Service," meaning reselling access to this model to clients, for example via a subscription or usage-based billing. The license itself is clear on this point: if a company and its subsidiaries together exceed $20 million in cumulative revenue over a rolling twelve-month period in this resale activity, they must sign a separate commercial agreement with Moonshot before continuing to use it.

In short: a bank that installs Kimi K3 internally for its customer service owes nothing to anyone. A large cloud provider reselling access to the model to thousands of clients, however, quickly crosses the threshold and must pay its dues.

A second rule adds to this, more visible to the general public: after reaching 100 million active users each month (or $20 million in monthly revenue), the name "Kimi K3" must be clearly displayed in the interface of the product using it. Thus, it is impossible to run the model behind the scenes under one’s own brand once a certain level of success is achieved.

Researcher Nathan Lambert, who has led open model projects himself, summarized the situation on X: this license draws inspiration from classic open licenses in its form but is clearly designed to prevent free commercial use at scale.

Moonshot isn’t inventing anything, but it is choosing its side.

This practice is not a first. Meta already imposes similar rules on its Llama models beyond a certain number of users. The difference with Moonshot: the thresholds trigger a payment obligation, not just a simple prohibition. Its Chinese competitor DeepSeek, on the other hand, has made the opposite choice with its latest model, distributed without any of these conditions.

This choice can be easily explained by the company’s finances. Moonshot is preparing for an IPO and aims for a valuation of around $50 billion, compared to about $4 billion by the end of 2025. Offering the largest model on the market for free costs much less than paying to attract users one by one, a calculation made even more profitable since the initial announcement of such a powerful Chinese model had already been enough, just days earlier, to shake Wall Street and the semiconductor industry. However, this calculated generosity has very precisely fixed limits, and the bill awaits those who turn it into a real business.