Kimi K3 and the revenue-share license

What was released

Moonshot put the weights of Kimi K3 on Hugging Face on July 27. The model card lists 2.8 trillion total parameters, 104 billion active, with 896 experts of which 16 are selected per token plus two shared. The 93 layers mix 69 Kimi Delta Attention layers with 24 gated multi-head latent attention layers, and the context window is 1,048,576 tokens. A 401 million parameter vision encoder handles images and video. Moonshot describes the combination of delta attention, attention residuals and its latent MoE framework as buying roughly 2.5 times the scaling efficiency of K2.

The reported scores are frontier-adjacent. The card lists 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1, 67.5 on DeepSWE and 91.2 on BrowseComp. Nathan Lambert's writeup places it second on the Vals AI index and third on Artificial Analysis's intelligence index, behind Claude Fable 5 and GPT-5.6 Sol Max, and first on the Frontend Code Arena. By his count it is the third most capable model in the world and the strongest open-weights model by a clear margin, which is why he calls the release an escalation.

What the license actually says

The terms are where the release gets interesting, and the reporting on them has been looser than the text. Wikipedia's entry on Moonshot describes a custom license requiring revenue sharing of up to 30 percent for services above 20 million dollars a year. We read the license file in the repository and could not find a percentage in it. What the text does say is that a licensee whose aggregate revenue with affiliates from offering the model as a service exceeds 20 million US dollars over any consecutive twelve months must enter a separate commercial agreement with Moonshot. The terms of that agreement are not in the public document.

So the structure is a two-tier license. Below the threshold, or for internal use, or for access through Moonshot's own products and certified partners, the grant is broad: use, modify, distribute, keep the copyright notice, comply with the law. Above the threshold, for the specific activity of selling inference or fine-tuning access through an API, you need a deal. There is a second clause in the same spirit that K2's modified MIT license also had: a product that exceeds 100 million monthly active users or 20 million dollars in monthly revenue must display the name Kimi K3 prominently in its interface. If the 30 percent figure is real, it lives in those private agreements, and we would treat it as a report about negotiations rather than a term of the license until someone publishes a signed copy.

Is it open?

By the Open Source Initiative's definition, no, and it was never going to be. A license that discriminates by field of endeavour or by the size of the licensee fails the definition, and both K2 and K3 do that. The more useful question is whether the release does the work that open weights are supposed to do, and there the answer is mostly yes. A researcher can download the weights, fine-tune them, publish results and redistribute derivatives. A startup can build a product on them. The only party the license reaches for is a company that has already cleared 20 million dollars selling the model itself.

That is a narrower restriction than some of the bespoke licenses attached to Western weight releases that we have been calling open for years, which have tended to reach into who may use the model and for what. Kimi's threshold is written in dollars of model-as-a-service revenue and touches nothing else. We do not love it, and we would rather every lab used Apache 2.0, but if we have been counting those releases as open weights then consistency requires counting K3 too, with the same asterisk.

The part that worries us more than the percentage is the private agreement. A public license, even a bad one, is a known quantity that lawyers can price. A clause that says you will negotiate with us above a line introduces uncertainty exactly at the point where a hoster has the most at stake, and uncertainty is what keeps large infrastructure providers from committing to a model. If Moonshot wants K3 served everywhere, publishing the standard commercial terms would cost them little.

The endorsement from above

The release did not happen in a vacuum. Lambert notes that Xi Jinping used the World AI Conference to commit China's AI ecosystem to open source and global diffusion. Read together with K3, the largest open-weights model ever released, and with Alibaba's announced 2.4 trillion parameter Qwen 3.8 also slated for open weights, this looks like state policy rather than a sequence of individual lab decisions. The Chinese government has apparently concluded that current frontier models are an acceptable thing to give away, and that diffusion of Chinese models is worth more to it than the capability lead it might keep by holding them back.

Lambert's framing of the trade is the clearest we have seen. Open frontier models are decelerationist for the profits of closed frontier labs and accelerationist for diffusion through the wider economy. Whether that is good depends on where you sit. It also lands while the Commerce Department and the White House are reported to be weighing restrictions on Chinese open models, and while the House Homeland Security Committee has subpoenaed DoorDash, Airbnb and Cursor over their use of Kimi models. A restriction would leave US companies unable to use a model that everyone else on earth can, which is the asymmetry Lambert warns about.

What we would want from the next one

Moonshot's own history suggests it can tighten or loosen the terms on the next release. K2 shipped under modified MIT a year ago. K3 shipped under a bespoke license with a negotiation clause. The direction is toward more control, which is what you would expect from a company that reached 200 million dollars of annual recurring revenue by April and a 35 billion dollar valuation this month. Open weights were a way to gain share. Now they are a way to keep it while charging the biggest customers.

For those of us who want to build on these models, the ask is simple. Publish the commercial terms. Keep the research and small-business grant as broad as it is. And release the base model alongside the post-trained one, because a base checkpoint is what lets the research community check the claims about scaling efficiency instead of taking the model card's word for it. K3 is a serious contribution to the open ecosystem. It would be a much bigger one with those three changes.