Kimi K3 Open Weights Shipped: What the Licence Says

The Kimi K3 open weights shipped on July 27, 2026, and the licence attached to them is the story. Moonshot AI did not publish K3 under MIT or Apache 2.0. It published the model under a bespoke document titled the Kimi K3 License, tagged on Hugging Face as license:other, carrying a revenue-triggered separate-agreement clause and a user-interface attribution mandate that most coverage of the release never quoted.

That distinction matters because “open weights” is doing a lot of work in the marketing copy. You can download the checkpoints, run them on your own hardware, fine-tune them, and quantize them. What you cannot do — not without a separate negotiated agreement — is resell K3 as a Model-as-a-Service business past a specific revenue line. And if your product crosses a specific user or revenue threshold, the licence says the words “Kimi K3” have to appear on your interface.

This guide reads the actual licence text clause by clause, puts it side by side with three other open-weight licence regimes, and reconciles what shipped against what was reported in the run-up. We also separate the two very different terabyte-scale numbers now in circulation — the repository download size and the resident memory footprint — because conflating them will wreck a hosting estimate.

The Hugging Face model card and the GitHub README agree on the headline specification: 2.8 trillion total parameters with 104 billion activated per token, drawn from 896 experts of which 16 are selected per token. The architecture runs 93 layers — 69 Kimi Delta Attention layers and 24 Gated MLA layers — inside what Moonshot calls a Stable LatentMoE framework, with a 160K vocabulary and a MoonViT-V2 vision encoder at 401 million parameters.

The context window is 1,048,576 tokens. Weights are natively quantized to MXFP4 with MXFP8 activations, applied via quantization-aware training from the supervised fine-tuning stage onward — Moonshot frames that choice as being for broad hardware compatibility rather than as a post-hoc compression step. Moonshot also states a roughly 2.5× improvement in overall scaling efficiency over Kimi K2, attributed to the Kimi Delta Attention, Attention Residuals, and Stable LatentMoE combination. That efficiency claim is vendor-stated and has not been independently replicated.

One inconsistency is worth naming before anyone builds a roadmap on it. The GitHub feature prose markets K3 as understanding text, images, and video within the same model. The formal Model Summary specification table on the same page lists only text and image modalities — there is no video row. Treat text and image as the confirmed surface and treat video as marketing copy until the spec table says otherwise. We covered the announced specification in our pre-release coverage of the K3 announcement, and the operational prep work in the readiness checklist we published ahead of the weight drop.

We pulled the LICENSE file directly from the Hugging Face repository rather than relying on any summary of it. The document is short, and three sections carry all of the operative weight: a Model-as-a-Service revenue trigger, a user-interface attribution trigger, and a pair of exemptions that between them cover a large share of real enterprise deployments.

Section 2 is the clause that separates this licence from a permissive one. It states that where a licensee or any of its affiliates operates a Model-as-a-Service business and the aggregate revenue of that group exceeds twenty million US dollars in total over any consecutive twelve months, the licensee must enter into a separate agreement with Moonshot AI before using the software or its derivative works for any commercial purpose. Note the structure: the threshold is on the group’s aggregate revenue, not on revenue attributable to K3.

Section 3, quoted above, is the attribution mandate. It is a genuinely high bar — a product would need more than 100 million monthly active users or more than $20 million in monthly revenue before it bites — but it is an obligation on the interface, not a footnote in a docs page, and it is the kind of term that brand-sensitive companies review carefully before they ship.

Section 4 supplies the exemptions, and it is the reason most readers of this post will not be affected. The obligations in Sections 2 and 3 do not apply to purely internal use that is never exposed to third parties, nor to use accessed through Moonshot’s own official products or its certified inference partners. If you are running K3 inside your company for your own staff, or consuming it through a certified partner endpoint, the restrictions are dormant.

Most “is it really open?” commentary treats licences qualitatively — permissive, restrictive, community. That is not useful to a procurement reviewer, who needs to know which specific number their company is above or below. So we pulled four licence texts and put them on the same axes. The table below is our own synthesis; every threshold in it comes from the licence document itself.

Read across the rows and a pattern shows up that the “open versus closed” framing hides completely. These four regimes restrict entirely different populations. Meta’s licence restricts consumer platforms at genuinely enormous scale and ignores revenue. MiniMax’s licence restricts everybody by default and only relaxes on notice. Moonshot’s licence restricts one specific business model — reselling inference — and leaves application builders essentially alone.

The thresholds also look deceptively similar and are not. MiniMax measures $20M over a year for authorization. Kimi K3 measures $20M over any consecutive twelve months for the separate-agreement trigger, and $20M per month for the attribution trigger. Three different obligations, three different measurement periods, one recurring number. Keeping each licence’s figures scoped to that licence is the entire discipline here — and it is the sort of detail that belongs in the same review file as procurement risk in open-weight sourcing decisions.

Pre-release coverage of K3 accumulated a set of claims that the live artifacts do not support. This is not unusual — a model announced weeks before its weights ship generates a lot of secondhand reporting — but several of those claims are now circulating as settled fact. Below, each one is set against what the primary source returns, retrieved on release day.

The licence row is the one that changes decisions. A team that budgeted for K3 on the assumption of MIT-style terms may have planned a hosted-inference product around it. Under the actual licence, that product is the one use case explicitly gated — and the gate is on group revenue, so a large company launching a small K3 reselling business is over the line from day one, while a startup doing the same thing is not.

Post-release coverage is now circulating two figures in the same neighbourhood — roughly 1.56TB and roughly 1.4TB — and using them interchangeably. They measure different things, they are produced by different methods, and they feed different budget lines. Confusing them will produce a hosting plan that is wrong in both directions at once.

The MXFP4 arithmetic is worth doing by hand once, because it generalizes. Four bits per parameter is half a byte. A 2.8-trillion parameter model therefore needs about 1.4 trillion bytes of resident weight memory before you allocate a single token of KV cache, an activation buffer, or any headroom for the 1M-token context this model advertises. Those additional allocations are where long-context serving actually gets expensive, and they scale with concurrency rather than with the weights.

There is a second signal in the artifact size. As of our retrieval on July 27, 2026, the repository showed 2,850 downloads against 5,637 likes — likes running roughly 2× ahead of downloads. That ratio is a reasonable proxy for how many people are interested in the release versus how many can realistically pull a 1.56TB artifact on release day. Both numbers were moving during the day we checked them, so treat them as a snapshot rather than a stable fact.

K3 scores 57 on the Artificial Analysis Intelligence Index, which makes it the leading open-weights model on that index as of July 27, 2026. On overall standing across all models, open and closed, it sits 4th — a position stated directly in Modal’s own release-day blog post and consistent with the current top of the leaderboard.

That ranking needs a date stamp every time it is used, because it has already moved. An Artificial Analysis article published on July 17, 2026 placed K3 at #3 overall. That article predates the launch of Claude Opus 5, and the ranking is stale as a result. If you see #3 cited for K3, check the publication date before you repeat it.

This is the most interesting and most misused pair of numbers around K3. Accuracy improved by roughly 13 percentage points generation over generation. The hallucination rate rose by roughly 12 percentage points over the same comparison. A model that answers more questions correctly while also asserting more things that are not true is a model that has become more willing to answer — and that trade is a deployment decision, not a defect.

Two caveats govern how precisely you can use those figures. First, the underlying benchmark run was published on July 17, 2026 — ten days before the weights shipped — and was necessarily measured against the hosted API version rather than the checkpoints you can now download. Same underlying model, different serving path, and a measurement gap worth naming whenever precision matters. Second, the blended overall index in that same analysis moved from +6 to +18 between the two generations, which is the direction Moonshot will quote and the accuracy-versus-hallucination split is the direction a risk reviewer should quote.

For any workload where an unverified assertion is expensive — regulated content, financial summarization, medical or legal drafting — the operational answer is not to pick a model on an index score. It is to run your own eval on your own prompts and to wire retrieval and citation checks into the pipeline. That discipline sits alongside the controls we mapped in our AI agent governance and compliance guide.

Moonshot’s published rate card for K3 is $3.00 per million input tokens on a cache miss, $0.30 per million cached input tokens — a 90% discount — and $15.00 per million output tokens. That pricing is flat across the entire 1,048,576-token window, with no tiering by context depth, which is unusual enough among long-context models to be worth stating explicitly. The same figures are mirrored on third-party model pages.

Whether that represents a price increase over the previous generation is being discussed in trade press, and the framing there is that Chinese frontier providers are no longer competing purely on rock-bottom rates. We are not printing the predecessor’s exact per-token figures here: the comparison we found is single-sourced to secondary coverage and we could not confirm it against a Moonshot-published price sheet for that model. The directional claim stands; the specific numbers should not be quoted as fact.

On that methodology K3 runs about 48% below Claude Opus 4.8 and about 10% below GPT-5.6 Sol per task — worth caveating with the same July 17, 2026 date stamp as the accuracy figures above, since the cost-per-task number depends on how many reasoning tokens a model spends, and that changes with serving configuration.

Output speed is where the numbers get genuinely misleading if you are not careful. A cross-provider benchmark puts K3 at 33.3 tokens per second. Modal, on its own infrastructure and using a custom speculator it describes as tuned to K3’s shape, reports 460 tokens per second on release day. Those two figures measure different things — a blended view across providers versus one vendor’s optimized deployment — and averaging them produces a number that describes nothing. Quote whichever is relevant to your deployment path, with its source attached.

Strip away the headline arguments about whether K3 is “really open” and the licence resolves into four situations. Find yours, and the answer is usually short.

The forward-looking read matters more than any single clause. Two years ago the open-weight question was binary: were the weights downloadable or not. That question is now largely settled in the affirmative across Chinese frontier labs, and the competitive pressure has moved to the licence text. Meta gates on user scale. MiniMax gates on commercial use itself. Moonshot gates on one business model. Each of those is a deliberate commercial choice dressed as a legal document, and each targets a different competitor rather than a different risk.

Expect the trend to continue in one direction: more open weights, more conditional licences, and thresholds calibrated so that the labs’ actual competitors — hyperscalers and inference resellers — are the ones who have to pick up the phone, while researchers and application developers stay free. For buyers, this means the licence file becomes a standing item in model evaluation rather than a formality, and the reasonable procedure is to read the LICENSE file at adoption and re-read it at every major version. If you want that evaluation run properly across a shortlist, our AI and digital transformation engagements start with exactly this kind of comparative review, and our web development team handles the integration work once a model clears it.