Open Weights With a License Attached Is Not Open Source. Kimi K3 Just Proved the Point.

On July 27, Moonshot released the full weights for Kimi K3. All 2.8 trillion parameters, 104 billion active. The benchmarks are real, the model is a monster, and the download is free. What it is not, despite what half the coverage says, is open source. K3 shipped under a custom license, and this time Moonshot did not even pretend otherwise. Simon Willison pointed out that the license text no longer calls itself modified MIT. It is just a license now, with terms, and the terms are the story.

A few weeks ago I wrote about the difference between open weights and open source, the Llama MAU cap, Gemma's usage restrictions, the OSI definition that almost none of these releases satisfy. That piece was the general principle. K3 is the case study, and it arrived faster than I expected.

What the license actually says

Two clauses matter, and VentureBeat published the relevant text. First, if you or your affiliates operate a Model-as-a-Service business and your aggregate revenue exceeds twenty million dollars over any consecutive twelve months, you must enter a separate agreement with Moonshot before any commercial use. Second, if the software or its derivatives power a commercial product with more than one hundred million monthly active users, or more than twenty million dollars in monthly revenue, the words Kimi K3 must be prominently displayed on your product's user interface.

Read that second one again. Not in your docs. Not in a model card. On your UI. Enterprises that abstract away the underlying model, which is most of them, now have a disclosure obligation that kicks in exactly when they reach scale. And the MaaS clause is not hypothetical for anyone. Together, Fireworks, and Baseten hit it on day one. Every serious inference provider now needs a bilateral deal with Moonshot to serve K3 commercially.

The play, and who invented it

Here is the sequence Moonshot ran. Release the weights, win the news cycle, top the leaderboards, sell out the API, and collect the goodwill. Twenty five American companies signed letters defending open weights with K3 as the poster child. Then the actual license landed, with revenue gates and mandatory branding.

This is not a Chinese invention. Llama pioneered the move, open weights with a commercial tripwire buried in the license, and it worked so well that it became the template. What K3 confirms is that open has become a release strategy rather than a commitment. The word does the marketing. The license does the business.

The gates are set lower than you think

The subtle part is the economics. Llama 4's 700 million MAU cap only ever binds giants. If you hit it, you are Snapchat or bigger, and you have lawyers. K3's gate is twenty million dollars of revenue over twelve months. That catches mid-size startups. That catches a Series A company with real traction. Moonshot wrote a license that converts successful startups into licensing customers, automatically, at the exact moment they become worth converting. It is a freemium funnel wearing an open-weights costume.

The branding clause is genuinely new territory. Mandatory display of Kimi K3 in your interface at scale is advertising extracted as a license term. Your product becomes Moonshot's billboard precisely when your product becomes worth looking at. No OSI-approvable license does anything like this, and no OSI-approvable license ever will, because the whole point of open source is that success does not trigger obligations to the vendor.

The fair counterpoint

Let me be honest about the other side. Moonshot spent an enormous amount of capital training K3, and it has released far more than OpenAI or Anthropic ever have. Wanting a revenue share from businesses built directly on top of your model is a reasonable commercial position. I do not begrudge them the terms.

The problem is not the terms. The problem is the vocabulary. You do not get to collect the reputational dividend of open source, the community goodwill, the policy letters, the leaderboard halo, while your legal text says something else. If the license has a revenue gate, call it a source-available commercial license and take your applause honestly. Closed APIs, whatever else you think of them, at least offer an honest deal. You pay, you get access, nobody pretends.

The taxonomy is now required diligence

The framework from my earlier post is no longer optional reading, it is the diligence lens. Tier one is true open source, MIT or Apache 2.0, irrevocable, no strings. DeepSeek, Qwen, Mistral's Apache releases, Pythia, OLMo. Tier two is conditional open weights, free until you succeed, at which point the vendor appears. Llama with its MAU cap, and now K3 with its revenue gate. Tier three is closed APIs, where the deal is at least stated up front.

For every model in your stack, know which tier it sits in and what happens to it when you win. That second question is the one people skip, because the license costs nothing today, and today is when the decision gets made.

What founders should actually do

First, K3 is genuinely useful and genuinely free below the gates. Prototype on it. Run internal workloads on it. The model is excellent and nothing in the license touches you at that stage.

Second, if your growth plan crosses twenty million dollars in revenue, and if it does not, why are you raising, then price in a Moonshot negotiation now or architect for a model swap later. Abstraction layers are cheap when you build them early and brutally expensive when you retrofit them under a compliance deadline.

Third, make the tier check a habit. Every model that enters your stack gets classified before it ships, not after your general counsel finds the clause during your Series B diligence. The defining feature of conditional open weights is that the license only binds when you win, which is exactly when you can least afford the surprise.

The future of open AI looks like this license. More weights will ship, more custom terms will ride along with them, and the word open will keep doing marketing work its legal text does not back. You cannot fix the vocabulary, but you can refuse to be fooled by it. Read the terms as if you will succeed, because that is when they activate.