--- title: "Kimi K3's Weights Are Open. Its License Isn't What Anyone Reported." date: 2026-07-28 tags: ["kimi-k3","moonshot-ai","open-weight-models","licensing","open-source","mixture-of-experts"] categories: ["AI Tools","Industry"] summary: "Moonshot AI shipped Kimi K3's full weights on July 27 — 2.8T total parameters, 104B active, 1M context, the largest open-weight release to date. But the license isn't the 'Modified MIT' terms that circulated before launch: it's a bespoke document with a $20M MaaS revenue gate and a mandatory UI-branding clause above 100M MAU, and several outlets reported the permissive version anyway." --- ![Kimi K3's Weights Are Open. Its License Isn't What Anyone Reported.](/images/kimi-k3-open-weights-license-surprise.png) Moonshot AI published the full weights for Kimi K3 on July 27 — a day ahead of its own July 27 target window closing, and by most measures the largest open-weight model release to date. 2.8 trillion total parameters, 104 billion active per token, a 1M-token context window, and day-zero inference support from both vLLM and SGLang. That part of the story is exactly what the pre-release coverage promised. The part that isn't: the license. For weeks, the model was reported as shipping under permissive, "Modified MIT"-style terms. The actual LICENSE file on Hugging Face says something considerably narrower — and at least one outlet ran the permissive version as a headline anyway, after the real terms were already public. ## What Actually Shipped Kimi K3 is a Mixture-of-Experts transformer with 896 experts, routing 16 per token plus 2 shared experts through what Moonshot calls its Stable LatentMoE framework, for roughly 104B active parameters out of 2.8T total. It uses a new attention mechanism, Kimi Delta Attention, which Moonshot claims makes long-context inference up to six times cheaper than prior approaches, and supports native multimodal input — text, image, and video — alongside the 1M-token window. The weights are published in MXFP4 (quantization-aware training carried through from the supervised fine-tuning stage), bringing the download to roughly 594GB versus 1.56TB at full precision — still a genuinely large model to self-host, realistically requiring an 8-node cluster of 8×80GB GPUs as a practical floor. Day-zero tooling support was real, not promised-later: vLLM shipped inference support at launch, reporting 118 tokens/second at baseline and up to 370 tokens/second with speculative decoding on a 16× GB300 NVL72 cluster. SGLang support landed the same day via LMSYS's Miles project, done in direct collaboration with Moonshot and Nvidia. On Moonshot's own self-reported benchmarks, K3 posts 88.3 on Terminal-Bench 2.1, 42.0 on SWE Marathon, and 77.8 on Program Bench — competitive framing against frontier proprietary models, though none of these numbers have an independent, controlled reproduction yet. Treat them the way you'd treat any vendor's own benchmark table: directionally interesting, not settled. ## The License Nobody Reported Correctly Here's the part that matters more than the parameter count. In the weeks before the weight drop, secondary coverage repeatedly described Kimi K3's license as a "Modified MIT" permissive grant — reasonable-sounding shorthand that then got repeated as settled fact across enough blogs and aggregators that it started to look confirmed. It wasn't. Hugging Face lists the license as `license:other`, and the actual document is a bespoke "Kimi K3 License" with two commercial gates that a stock MIT or Apache license simply doesn't have: - **A Model-as-a-Service revenue trigger.** If you operate a MaaS business — defined in the license as one where third parties can exercise "meaningful control" over inputs, parameters, or training data — and your aggregate revenue from that business exceeds $20 million over any trailing 12 months, you need a separate commercial agreement with Moonshot before you can keep operating on K3. Internal deployments and products that simply embed the model as a feature are explicitly exempted; the clause is aimed squarely at businesses reselling inference access to K3 itself. - **A mandatory UI-branding requirement.** Any product or service crossing 100 million monthly active users, or $20 million in monthly revenue, has to display "Kimi K3" prominently in its user interface. That's not a request in a README — it's a license term. This isn't unprecedented as a category — Meta's Llama license has long gated by user scale, and plenty of "open" model releases carry some kind of commercial asterisk. What's notable here is the gap between what got reported and what's actually in the file. At least one outlet (AIToolsRecap) published a piece asserting the "Modified MIT License Confirmed" days after the real, restrictive license text was already sitting on Hugging Face for anyone to read. That's not a subtle documentation nuance getting flattened in summary — it's the opposite claim, stated as fact, after the correction was trivially available. ## Why This Is the Real Story If you're evaluating Kimi K3 to self-host inside a commercial product — and a 2.8T open-weight frontier-class model with day-zero vLLM/SGLang support is a legitimate thing to evaluate — the license is not a footnote. It's the difference between "we can build a business on this" and "we need Moonshot's sign-off first." The MaaS clause specifically targets the exact business model most likely to get built on top of a newly open-weighted frontier model: standing up hosted inference and reselling access. If that's your plan, "Modified MIT" and "$20M MaaS revenue trigger requiring a separate agreement" are not interchangeable descriptions, and building a go-to-market plan on the wrong one is an expensive way to find out. There's a broader pattern worth naming here too. Every major open-weight release this year has arrived wrapped in the same reflexive "it's basically open source" framing, regardless of what the license actually says — because "open weights, permissive license, largest model ever" is a more shareable headline than "open weights, license TBD, here's what to check before you build on it." That gap between hype-cycle summary and primary-source license text is exactly the kind of due-diligence failure that costs teams real money when it surfaces after the fact, not before. If you're the one making the call on whether to self-host a model like this inside your product, read the LICENSE file yourself before you read anyone's summary of it — including this one. It's also a useful contrast for anyone weighing "open weight" against Claude's straightforwardly proprietary, API-only model. Anthropic doesn't pretend Claude is open, and there's no LICENSE-file surprise waiting for you six weeks after you've built a product on it — the commercial terms are the commercial terms, published upfront. "Open" that turns out to have a revenue-triggered commercial gate isn't more honest than "closed" that says so from the start; it's just a different kind of fine print, discovered later. --- **Sources:** - [Hugging Face: Kimi K3 Model Overview — MXFP4 Quantization and What the Open Weights Mean](https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei) - [Digital Applied: Kimi K3 Open Weights Shipped — What the Licence Says](https://www.digitalapplied.com/blog/kimi-k3-open-weights-shipped-license-restrictions-2026) - [Trending Topics: Kimi K3 — New License Makes Clear How Moonshot AI Wants to Make Money](https://www.trendingtopics.eu/kimi-k3-new-license-makes-clear-how-moonshot-ai-wants-to-make-money/) - [vLLM Blog: Day-0 Kimi K3 Inference Support](https://vllm.ai/blog/2026-07-27-k3) - [LMSYS/SGLang Blog: Kimi K3 Day-0 Support via Miles](https://www.lmsys.org/blog/2026-07-27-kimi-k3-day0-support)