Kimi K3 License Explained: Commercial Use, the $20M Rule & Limits
September 25, 2026 · kimi k3 license · kimi k3 commercial use · kimi k3 open source · open weight license · model as a service · moonshot ai · ai model licensing
Yes, you can use Kimi K3 commercially for free, and for most companies the license behaves like MIT. The Kimi K3 License grants the right to use, copy, modify, fine-tune, distribute and sell the model and its weights without payment. It adds two conditions that only bite at scale: a company running a Model-as-a-Service (MaaS) business with more than $20 million in group revenue over any 12 months must sign a separate agreement with Moonshot AI first, and any product with more than 100 million monthly active users or $20 million in monthly revenue must display “Kimi K3” prominently in its interface. Purely internal use is exempt from both.
This breakdown is based on the actual LICENSE file in the official Hugging Face repository and its identical copy on GitHub, checked on September 25, 2026. Commit history shows the file has not changed since the initial commit on July 27, 2026, the day the weights went public.
Not legal advice. This is an informed plain-English reading by a fan site, not a lawyer. If real money or a large user base depends on your interpretation, email Moonshot at [email protected] (the contact the license itself gives) and have counsel review it.
The Kimi K3 License at a glance
Five numbered sections, about 3,000 characters in total.
| Section | What it says | Who it actually affects |
|---|---|---|
| Grant | Free rights to use, copy, modify, merge, publish, distribute, sublicense and sell; to run, deploy, fine-tune and create derivative works. Explicitly covers weights, parameters, configs, inference and training code | Everyone |
| 1. Notice | Keep the copyright notice and license text in all copies; comply with applicable laws | Anyone who redistributes |
| 2. Model as a Service | MaaS operators whose licensee + affiliates revenue tops $20M over any 12 consecutive months need a separate agreement with Moonshot before commercial use | Inference clouds, API resellers, fine-tuning platforms, big-company subsidiaries |
| 3. Attribution | Commercial products with >100M MAU or >$20M monthly revenue must prominently show “Kimi K3” in the UI | Very large consumer or enterprise products |
| 4. Exemptions | Sections 2 and 3 do not apply to internal use, or to use through Moonshot’s official products or certified inference partners | Enterprises self-hosting internally; customers of certified hosts |
| 5. Warranty | Software “and any output and results therefrom” provided as-is; no liability for Moonshot | Everyone |
Hugging Face tags it license: other (license_name: "kimi-k3"). That is accurate: the text is MIT-shaped, but it is not MIT, and not the Modified MIT of the K2 generation either.
Clause by clause: what the text actually says
The grant is broader and more explicit than MIT
Classic MIT covers “this software and associated documentation files.” The K3 License spells out that the “Software” includes “the model weights, parameters, configuration files, inference and training code,” and adds the rights “to run, deploy, fine-tune, or otherwise modify the Software and create derivative works from it.” That removes a K2-era ambiguity about whether an MIT-style software license cleanly covers weights. Fine-tuning and derivatives are now named rights.
Section 2: the Model-as-a-Service gate
The clause that makes K3 different. The definition, verbatim:
“Model as a Service” means giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data.
Two things are carved out: “(a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.”
The trigger then reads: if the licensee “or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars … in total over any consecutive 12 months,” a separate agreement with Moonshot is required “before using the Software or its derivative works for any commercial purpose.”
Three details matter more than most coverage admits:
- It measures total group revenue, not K3 revenue. A startup API host making $5M a year is under the line. A subsidiary of any sizeable company crosses it on day one, even if its K3 endpoint has earned nothing (Digital Applied and Implicator both flag this).
- It is a rolling 12-month window, so a fast-growing host can cross it mid-year.
- It is prior permission. The wording is “before using,” not “within X days.”
Section 3: the attribution rule (inherited from K2)
If the Software or any derivative “is used for any of the Licensee’s commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars … in monthly revenue, ‘Kimi K3’ must be prominently displayed on the user interface.” This is the same sentence the K2 licenses used, with the model name swapped.
Note the asymmetry: the MaaS trigger is $20M per year of company revenue, while the attribution trigger is $20M per month from the specific product, or 100M MAU. The first is a low bar aimed at hosts; the second only touches very large products.
Section 4: the two exemptions
Sections 2 and 3 do not apply to:
- Internal use, defined as “any use that does not make the Software, its outputs, or its underlying capabilities available to third parties.” Outputs count: shipping K3-generated answers to customers is not internal use.
- Use “accessed through Moonshot AI’s official products or certified inference partners.” If you consume K3 through Moonshot’s API or a certified host, you are not the one who has to negotiate. Moonshot has not published a list of certified partners, a gap Trending Topics also noted.
Section 5: warranty, and a quiet mention of outputs
The disclaimer covers “the Software and any output and results therefrom.” Responsibility for what K3 generates and you publish sits with you.
What changed from the Kimi K2 license
We pulled the LICENSE from every K2-series repo. All four are titled “Modified MIT License” with a single addition (“Our only modification part is that…”) containing the attribution rule.
| Kimi K2 / K2.5 / K2.6 / K2.7-Code | Kimi K3 | |
|---|---|---|
| License name | Modified MIT License | Kimi K3 License |
| Weights explicitly covered | No (generic “software”) | Yes, named in the grant |
| Fine-tuning / derivatives named | No (implied by “modify”) | Yes, explicitly |
| Attribution rule | >100M MAU or >$20M monthly revenue | Same |
| MaaS revenue gate | None | $20M group revenue / 12 months |
| Internal-use exemption | Not needed | Yes, for Sections 2-3 |
| Certified-partner exemption | None | Yes |
| Outputs mentioned | No | In warranty and internal-use definition |
Sources: K2-Instruct, K2.5, K2.6, K2.7-Code. If you plan to upgrade from K2 to K3, attribution is unchanged; the only new question is whether you run a MaaS business.
What counts as Model as a Service? The gray zones
Clear at the edges, fuzzy in the middle.
Clearly MaaS: a public API where customers send arbitrary prompts and set parameters; a hosted fine-tuning service where customers upload training data.
Clearly not MaaS: a SaaS feature using K3 behind the scenes (ticket summaries, product descriptions, classification) where the customer never touches the prompt. A gateway like OpenRouter forwarding requests to other companies’ K3 deployments is “mere relaying.”
Genuinely unclear:
- General chat assistants. Users typing anything into your chatbot arguably have “meaningful control over the inputs.” But it is also an “end-user product,” and the carve-out’s mention of “harnesses” suggests agent wrappers are meant to be exempt.
- Coding agents. An agent with its own system prompt and tool loop looks like a harness; a thin pass-through wrapper looks like an API.
If your company’s total revenue is under $20M a year, none of this matters yet. If you are above it and in the gray zone, send exactly this question to [email protected].
Fine-tuning and redistribution
Fine-tuning is explicitly allowed, and so is publishing the result. The rules for a derivative:
- Include the notice. Ship the LICENSE file and copyright notice with your weights (Section 1).
- The conditions travel with it. Sections 2 and 3 apply to “the Software or its derivative works,” so a published K3 fine-tune still carries the MaaS gate and attribution rule for whoever deploys it at scale.
- No naming rule. Unlike Meta’s Llama 4 license, which requires “Llama” at the beginning of any derived model’s name, K3 lets you call your fine-tune whatever you like.
- Fine-tuning as a service is MaaS. The definition explicitly includes “fine-tuning” access where customers control “training data.” A platform that sells K3 fine-tunes to others is squarely in Section 2.
Practically, the checkpoint is roughly 1.56TB of MXFP4 weights, so this is a data-center project; see our VRAM requirements guide.
What about outputs and distillation?
The weights license does not restrict K3’s outputs: no acceptable-use policy, no ban on training other models on its generations, nothing beyond “comply with applicable laws and regulations.” Llama 4, by contrast, incorporates an acceptable-use policy and requires “Llama” naming if outputs train a distributed model; MiniMax M3 appends a prohibited-uses list.
The API is a different contract. If you call K3 through Moonshot’s platform rather than self-hosting, the Kimi OpenPlatform Terms of Service (last updated July 30, 2026) apply. They prohibit using the service for “developing, serving, or creating applications, products, Services, or models that have potential competitive possibilities with the Services without authorization,” and they state that customer content may be used to improve Moonshot’s services unless otherwise agreed in writing. So: distilling from weights you host yourself is not restricted by the license; distilling from the official API runs into the ToS.
How K3 compares with other open-weight licenses
Chinese frontier labs have converged on one idea in 2026: keep the weights free, gate the hosting business. Z.ai’s GLM-5.3 license uses a Model-as-a-Service definition that is nearly word-for-word identical to Moonshot’s, and Alibaba’s flagship Qwen3.8-Max license uses a close variant. Their smaller siblings stay permissive.
| Model (license) | Commercial use | Big-company gate | Attribution | Use restrictions |
|---|---|---|---|---|
| Kimi K3 (Kimi K3 License) | Yes, free | MaaS + >$20M group revenue/12 mo: separate agreement | ”Kimi K3” in UI at >100M MAU or >$20M/mo | Applicable law only |
| Kimi K2.x (Modified MIT) | Yes, free | None | ”Kimi K2.x” in UI at >100M MAU or >$20M/mo | None |
| DeepSeek V4.1 Flash / V4 Pro (MIT) | Yes, free | None | Keep notice only | None |
| GLM-5.3 (GLM-5.3 License) | Yes, free | MaaS + >$10B group revenue/12 mo: pass Z.ai security review | None beyond notice | Applicable law |
| GLM-5.3 Flash, GLM-5.2 (MIT) | Yes, free | None | Keep notice only | None |
| Qwen3.8-Max 2.4T (Qwen3.8-Max License) | Yes, free | MaaS or “AI Work Assistant” + >$50M group revenue/12 mo: separate license | Model name in UI at >100M MAU or >$20M/mo | Law, no third-party IP infringement |
| Qwen3.8-27B (Apache 2.0) | Yes, free | None | NOTICE file | None (patent grant included) |
| Llama 4 (Community License) | Yes, free | >700M MAU on release date: must request license | ”Built with Llama”; “Llama” prefix on derived models | Acceptable Use Policy |
| MiniMax M3 (Community License) | Only with conditions | >$20M yearly product revenue: prior written authorization | ”Built with MiniMax M3” for any commercial use | Prohibited-uses appendix |
A few takeaways:
- K3’s MaaS bar is the lowest. $20M versus Qwen3.8-Max’s $50M and GLM-5.3’s $10 billion. Moonshot is aiming at mid-sized inference clouds, not only hyperscalers.
- Llama gates on users, Moonshot on business model. Meta’s 700M MAU test is measured once, on the release date, and hits a handful of giants.
- For product builders, K3 is looser than Llama: no badge below 100M MAU, no naming rule, no acceptable-use policy.
For head-to-head model comparisons beyond licensing, see Kimi K3 vs DeepSeek V4.1 Flash and Kimi K3 vs GLM-5.3 Flash.
Why Moonshot added the MaaS clause
Under the K2 licenses, any cloud could serve Moonshot’s model and keep all the revenue. In August, The Next Web summarized a Reuters report that Moonshot was in early talks with Microsoft, Amazon and Google, seeking up to 30% of the revenue they earn from K3, and stressed that “nothing has been agreed.”
On September 18, 2026, AWS announced Kimi K3 is generally available on Amazon Bedrock. The South China Morning Post called it a test of whether open-weight Chinese AI can earn hosting revenue; financial terms were not disclosed. AWS obviously clears the $20M bar, so an agreement presumably exists.
A correction: some summaries, including Wikipedia at the time of writing, say the license itself requires “revenue sharing of up to 30%.” It does not. The text names no percentage; the 30% is a reported negotiating position.
Which scenario are you? Four common cases
| You are… | Section 2 (MaaS) | Section 3 (attribution) | What to do |
|---|---|---|---|
| Startup SaaS embedding K3 in features | Not MaaS; under $20M anyway | Only above 100M MAU / $20M monthly | Keep the notice if you ship weights; otherwise just build |
| Enterprise self-hosting for employees | Exempt (internal use) | Exempt (internal use) | Comply with law; revisit if outputs reach customers |
| Lab fine-tuning and publishing weights | Not MaaS by itself | Applies to whoever deploys at scale | Ship the LICENSE file; conditions pass downstream |
| API reseller / inference host | Yes, if group revenue >$20M/12 mo | Possibly, at huge scale | Get a Moonshot agreement or certification before commercial use |
Startup SaaS. If K3 powers features inside your product (a writing assistant, a support bot with a fixed harness), you are almost certainly outside the MaaS definition. Calling K3 through Moonshot’s API or a host from our API providers guide also shifts the MaaS question onto them. Watch out for affiliates: a corporate parent’s revenue counts if you ever launch an API product.
Enterprise internal use. Employee-only deployments (internal search, code review, document analysis) are exempt from Sections 2 and 3 regardless of company size. The trap is scope creep: once K3 outputs reach customers, you are no longer “internal” and must re-check both tests.
Fine-tune and redistribute. Publishing a fine-tune does not by itself make you a MaaS operator. Offering it as a paid API does.
API reseller. This is who the license targets. Under $20M trailing group revenue, host freely. Over it, get an agreement or certified-partner status first, and ask Moonshot directly rather than assuming you qualify.
Is Kimi K3 “open source”?
Not by the formal definition. The OSI’s Open Source AI Definition 1.0 requires freedom to “use the system for any purpose and without having to ask for permission.” Section 2 requires certain businesses to ask permission, so K3 fails that test. Training data is not released either. The accurate term is open-weight: the full 2.8T-parameter checkpoint (104B active) is downloadable and commercially usable by nearly everyone, with conditions. For the release details, see our Kimi K3 open weights breakdown.
For most developers it is still more permissive than Llama or MiniMax.
Honest caveats
- Not legal advice. Courts and Moonshot may read the text differently than we do.
- Key terms are undefined: “meaningful control,” “prominently displayed,” “certified inference partners,” and how “monthly revenue” is attributed to a product.
- No governing law or venue is specified.
- Licenses are per release. A future K3.x or K4 could ship under different terms, as K3 did relative to K2.6.
- Hosted access adds another contract: Moonshot’s platform ToS or your provider’s terms.
FAQ
Can I use Kimi K3 commercially for free? Yes, for most businesses. The Kimi K3 License lets anyone use, modify, fine-tune, redistribute and sell the model free of charge. Only two groups get extra obligations: companies running a Model-as-a-Service business whose group revenue exceeds $20 million over any 12 months (they need a separate agreement with Moonshot), and products with more than 100 million monthly active users or $20 million in monthly revenue (they must display “Kimi K3” in the UI).
Is the Kimi K3 License the same as the Kimi K2 modified MIT license? No. Every K2-series release (K2, K2.5, K2.6, K2.7-Code) used a Modified MIT License whose only addition was the 100M MAU / $20M monthly revenue attribution rule. The Kimi K3 License keeps that rule but adds a new Model-as-a-Service clause with a $20 million aggregate revenue trigger, plus exemptions for internal use and for access through Moonshot’s official products or certified inference partners.
What counts as Model as a Service under the Kimi K3 License? The license defines it as giving a third party access to language model inference or fine-tuning (for example via an API) in a way that lets them exercise meaningful control over the inputs, parameters or training data. It excludes end-user products where the model is solely embedded within specific features or harnesses, and the mere relaying of requests to models hosted by others.
Can I fine-tune Kimi K3 and publish my own model? Yes. The license explicitly grants the right to fine-tune, modify and create derivative works, and to distribute and sublicense them. You must include Moonshot’s copyright notice and the license text, and your derivative carries the same Section 2 and 3 conditions. Unlike Llama, there is no requirement to put the original model name in your derivative’s name.
Does the Kimi K3 License restrict how I use the model’s outputs? The weights license places no restrictions on outputs and has no acceptable-use policy beyond complying with applicable law. If you use Moonshot’s hosted API instead, the Kimi OpenPlatform Terms of Service apply, and those prohibit using the service to develop competing models without authorization.
Is Kimi K3 open source? It is open-weight, not open source in the OSI sense. The OSI’s Open Source AI Definition 1.0 requires that a system can be used for any purpose without asking permission, and the K3 license requires some large Model-as-a-Service operators to sign a separate agreement first. The training data is also not released.
Does Moonshot take a 30% revenue share from Kimi K3 hosts? Not under the license text. The license only says large MaaS operators must enter a separate agreement; it names no percentage. The 30% figure comes from an August 2026 Reuters report that Moonshot was seeking up to 30% of K3 revenue in talks with Microsoft, Amazon and Google. Actual terms, including for Amazon Bedrock, have not been disclosed.
The bottom line
The Kimi K3 License is MIT for builders and a toll booth for large hosts. If you are embedding K3 in a product, running it internally, or fine-tuning and publishing weights, you can proceed today at no cost: keep the notice, and remember the “Kimi K3” badge only kicks in at 100M users or $20M a month. If you sell K3 inference or fine-tuning to others and your group revenue is above $20M a year, you need Moonshot’s signature first. That clause is the real change from K2, and it is how Moonshot plans to earn money from the hyperscalers now hosting its model.
Keep reading: the Kimi K3 open weights release, every Kimi K3 API provider, Kimi K3 VRAM requirements, Kimi K3 vs Kimi K2, and the complete Kimi K3 guide.