The Chinese company Moonshot AI released the weights for its much-discussed new Kimi K3 large language model today, as it promised when the model debuted exclusively through their platform on July 16. This means anyone with the (considerable) requisite computing resources can now download and run the model without interacting with its maker. The 1.56TB package is now available on Hugging Face.
The license Moonshot attached to the model restricts one key type of use, however. If you wish to provide the Kimi K3 model as a cloud service to resell, and if you have revenue over $20 million USD per year, you’ll need to “enter into a separate agreement with Moonshot AI.” This $20 million annual revenue threshold is comically low; almost all cloud platforms that resell models already trigger this condition. In effect, any inference provider—from Amazon Web Services down to GPU-first “neoclouds”—will have to enter into a commercial relationship with Moonshot to serve up Kimi K3. It’s a good bet that Moonshot will get some revenue sharing out of these deals. Companies or researchers that want to use Kimi K3 internally should still be able to do so without paying a dime. But if they want to sell the model as a service, they’ll have to make a deal.
Given how expensive it is to train a model of Kimi K3’s size and capabilities, attempting to make some money is understandable, as Moonshot stares down a not-so-distant future of becoming a public company. However, at a time when the US government is debating various ways to potentially block or discourage US use of Chinese-built open-weights models, this desire for a new revenue source also introduces a new source of risk.
MAKING REVENUE SHARING EXPLICIT
Kimi’s new[1] license provision looks like an attempt to walk a tightrope. Engineering and training large language models is expensive. The natural (though far from proven) business model to recoup those costs and hopefully make a profit is to sell access to them. Yet Chinese labs have released many of their most capable models openly, allowing companies and well-resourced individuals to run them for free. The only way they’ve been making money so far is serving their models behind their own APIs, which has limited reach. This means other companies are capturing some of the revenue that the open model makers might theoretically have collected.
On the other hand, openness is good for visibility and feedback. For instance, practically no one in the United States had heard of Moonshot AI, founded in 2023, until a gradually growing cohort of users found earlier Kimi models effective and cost-efficient for coding tasks over the last year or so. The first generation Kimi model, Kimi 1, was actually a closed model. The reaction and traction was so poor that the company decided to open weights for future generations. This month, with that growing visibility, Kimi’s K3 release landed on the front page of the New York Times.
To the extent that prominence drives direct business and investor interest, openness helped Moonshot level up quickly. Moreover, Moonshot has limited computational resources to serve model users; it even had to pause new sign-ups amidst the rush of attention around the K3 launch. And some users will not want to send their sensitive data outside their own company’s trusted computing infrastructure, let alone to systems controlled by a Chinese company.
So while China’s cloud-computing leader AliCloud has explored a business model where the most advanced version of their Qwen model was available exclusively from Alibaba’s cloud, hopefully catching some revenue and getting users used to their platform, Moonshot has tried something different. The company almost surely needs other providers to meet demand, and they can cut deals to bring home some of the revenue.
CHANGING THE LANDSCAPE OF REGULATORY RISK
This sounds like a good way to retain the benefits of open-weights release while pulling in some additional revenue to make the whole business more sustainable. Yet this license changes the nature of Moonshot AI’s legal relationship with inference providers around the world.
Previously, the providers could download the model, get it running on their infrastructure, and start selling tokens. No need for any explicit deal with the Chinese firm behind the model weights. Peter Harrell, a former Biden administration official and an expert in export and transaction controls, has argued that it’s unclear whether courts would allow the government to prohibit this pure form of open-weights model use. Considering various potential policy tools, he suggests that downloading the open models could be deemed an exempt transfer of “information,” rather than a controlled “service” or other transaction.
But if a US company needs a contract with Moonshot to provide the inference tokens that Kimi K3 generates, the picture looks different. Some of the policy tools US officials and others have debated as potential levers to restrict Chinese open model use would more clearly apply. For example, if the government decided to use the so-called ICTS rule, thus far a little-used mechanism by which the Department of Commerce can block “transactions” with “foreign adversary” entities that it deems to pose national security risks, it’s harder to argue there’s no transacting going on. If they put Moonshot on a sanctions list of entities US persons cannot legally do business with, then the contract to provide inference is clearly an instance of doing business. The commercial agreements this new license is asking for (and getting) from US companies will avail Moonshot to American jurisprudence on a whole new level.
This isn’t an argument about whether such restrictions are a good idea, but an analysis of the policy implications this new license introduces. As Moonshot seeks to generate more revenue from its work while still holding onto the halo effect of being “open,” it seems to have walked out onto a tightrope with a new balancing act: With revenue comes regulability.
[1] Kimi models previously included license provisions that required display of the model name when providing services above certain revenue or user thresholds, and similar language remains. The requirement for inference providers to make a deal is new.