The Largest Open-Weight Model Ever Isn't Open Source

On July 27, Moonshot AI put the full weights of Kimi K3 on Hugging Face. At 2.8 trillion parameters, it is the largest open-weight model anyone has ever released, and the download costs nothing. Most of the coverage stopped there, and several outlets reported it as shipping under the Modified MIT terms Moonshot used for the K2 series.

It did not. The repository is tagged license:other, and the document it points to is a bespoke instrument called the Kimi K3 License. That substitution is the most consequential thing about this release, and it is the part that got waved through.

What is actually in the file

Start with the artefact, because it is genuinely impressive. K3 is a mixture-of-experts model with 104 billion parameters activated per token, drawn from 896 experts with 16 selected on each forward pass. Context length is 1,048,576 tokens. Vision comes from a MoonViT-V2 encoder at 401 million parameters. The weights are natively MXFP4, with MXFP8 activations, applied through quantisation-aware training rather than bolted on afterward — which is why the numbers hold up rather than degrading the way post-hoc four-bit conversions usually do.

The download is roughly 1.56 terabytes across 118 files, resident at about 1.4TB once loaded. Running it means something on the order of eighteen 80GB accelerators, or a single node carrying eight 192GB cards — Blackwell-class Nvidia silicon or AMD's MI400. On third-party evaluations it lands as the strongest open-weights model available, taking first place on frontend-coding benchmarks and sitting fourth overall in late July, a placement that moved when Anthropic shipped Claude Opus 5 three days earlier.

So: a frontier-adjacent model, free to download, that requires several hundred thousand dollars of hardware to serve. That gap is where the licence does its work.

The two clauses

The Kimi K3 License is not a general-purpose restriction. It is aimed with some precision.

The first clause is a revenue trigger. Any operator running K3 as a model-as-a-service whose group aggregate revenue exceeds $20 million over any consecutive twelve months must negotiate a separate agreement with Moonshot before commercial use. Note the construction: the threshold is total group revenue, not revenue earned from K3. A profitable infrastructure company that wants to host these weights is over the line on day one regardless of whether a single customer ever calls the endpoint.

The second clause is an attribution mandate. Any product exceeding 100 million monthly active users or $20 million in monthly revenue must prominently display Kimi K3 in its user interface. Not in a model card. Not in documentation. On the surface the user looks at.

Section 4 then carves out the cases that would have generated the loudest complaints. Purely internal deployment, never exposed to third parties, is exempt. So is access through Moonshot's own products or its certified inference partners. Most enterprise use falls inside those carve-outs, which is precisely why the licence has attracted so little friction.

Read the exemptions to find the target

Once you line the triggers up against the exemptions, the intended defendant is obvious, and it is not the enterprise.

A bank fine-tuning K3 behind its own firewall is exempt. A developer running it on a workstation is irrelevant to the terms. The entity that the revenue trigger catches is the neocloud — the Together, Fireworks, Baseten tier of the market whose entire business is taking open weights and selling tokens off them. Those companies are the reason "open weights" became a commodity input in the first place. Under this licence they cannot serve K3 at scale without a negotiated deal, which means Moonshot sets the floor price on its own model rather than watching a dozen hosts race it to zero.

The attribution clause solves the other half. Moonshot's problem is not capability; it is distribution outside China. A logo requirement at 100M MAU is a customer-acquisition channel bought with parameters instead of ad spend — the model becomes free advertising the moment it succeeds inside somebody else's product.

The context makes the strategy sharper still. A week earlier, on July 21, Moonshot had to pause new subscriptions to its hosted Kimi service because demand overwhelmed its capacity to serve it. A lab that cannot buy enough compute to meet demand has an obvious move available: let other people's GPUs carry the inference load, and use the licence to retain the economics and the branding that hosting would otherwise have given away. Giving away the weights is not charity when serving them was the bottleneck.

Why the wrong label matters

None of this makes the release bad. Genuinely open weights at this scale would be a gift few companies can afford to make, and the engineering is real. But "open source" is a term with a definition, the Kimi K3 License does not satisfy it, and a licence that discriminates against a field of endeavour and imposes downstream display obligations is a commercial agreement wearing an open-source costume.

The distinction has teeth for anyone building on it. Under Modified MIT — the terms K2, K2.5, K2.6 and K2.7 shipped with — you could assume your obligations ended at the copyright notice. Under K3 you inherit a revenue test computed on your whole corporate group and a UI requirement that lands exactly when your product starts working. Teams that upgraded on the assumption of continuity did not read a new document, because nothing in the announcement suggested there was one.

The wider pattern is the thing to track. Chinese labs have spent two years using permissive licences to buy mindshare in a market where they could not compete on distribution. K3 is the first release at the frontier of that strategy where the licence tightens as the model gets good enough to matter. Expect the next one to tighten further. The weights were never the product. The terms attached to them are.