Open-Weight Licences Compared: Beam, Mistral, Muse, Kimi, Qwen

Yes, for most uses. On the licence files we read, Beam and Qwen3.8-27B (both Apache 2.0) allow commercial use, hosting and fine-tuning with no hosting condition. Mistral Large 4, Muse Spark Open, Kimi K3.1 Pro and the larger Qwen models also allow commercial use. They add conditions for businesses that sell model access to others, and some add naming rules. Read the file for your exact model. This is not legal advice.

Four large open-weight models shipped between mid-October and early November 2026. "Modified MIT" sounds harmless and "limited commercial terms" sounds vague. The only way to know is to open the LICENSE file, so we did.

Everything here was read on 9 November 2026. We quote only clauses we read. Where a page refused to load or a clause is unclear, we say so.

Can you use open-weight models commercially?

Commercial use is permitted subject to each licence's conditions. Internal use has exemptions, but products built on top of a model can still trigger restrictions. Qwen3.8-Flash-Next requires a separate licence for AI Work Assistant businesses, including standalone coding or office assistants, without a revenue threshold.

Start with the plainest case. Reflection AI released Beam under Apache 2.0, and the LICENSE file in its repository is titled "Apache License, Version 2.0, January 2004". Qwen3.8-27B also ships as Apache 2.0, and its repository file is the standard Apache text, with a 2026 copyright line filled in. The Apache text grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license", and its text never mentions hosting or running software as a service for others. We did not diff either file against apache.org, so treat that as a reading.

The other releases are custom licences built on the MIT idea. Each opens with a broad grant: use, copy, modify, distribute, sublicense, sell, run, deploy and fine-tune. Each then adds conditions. Those conditions fall into three kinds:

1. A hosting condition. If you run a Model-as-a-Service business, you need a separate agreement. Some licences set a revenue threshold and some do not.
2. A naming rule. At a stated scale, or for any product, you must show the model name.
3. A usage policy. Meta's licence attaches an acceptable-use policy.

The next section covers the definitions. Then comes one table.

What is the difference between open source and open weight?

Open weight means the trained parameters are downloadable. Open source, in the strict sense, means a defined set of freedoms that go beyond the weights. The two terms are often used as if they were the same, and they are not.

Two definitions matter here, and both come from the Open Source Initiative (OSI).

The Open Source Definition, version 1.9, is about software licences. Criterion 6 says "The license must not restrict anyone from making use of the program in a specific field of endeavor."

The Open Source AI Definition, version 1.0, goes further for models. It lists four freedoms: use, study, modify and share. It says users need "the preferred form to make modifications to the system". For machine-learning systems, that means data information, the complete code used to train and run the system, and the parameters. It also says releases only count as open-source weights if they "must include the data information and code used to derive those parameters."

Apply that to the table below. Apache 2.0 is an OSI-approved software licence. A model released under it, such as Beam, is open-weight with an open-source licence. We found no statement on Reflection's pages that the training data or full training code is published, so we do not call it open source AI.

The custom licences are a different case. They contain a condition aimed at one kind of business, which sits uneasily with criterion 6 of the Open Source Definition. We did not check whether OSI has reviewed any of them, and OSI decides what it approves, not us. The practical point is simpler. If a licence has a custom name, read it. Our earlier comparison of open and proprietary models made the same point for governance, and our model supply chain guide covers how to record licences next to file hashes.

How do the licences compare?

Here is what the files say. "Read" means we opened the licence file on Hugging Face. Mistral's file loaded, but the tool we used returned its clauses as short quotes and paraphrase, so we quote only short fragments for it.

Sources: the LICENSE or LICENSE.md file in each Hugging Face repository, plus the Apache text for the two Apache rows. Meta's own post calls its licence "the Muse Model License with limited commercial terms", which is a summary of the file.

Muse Spark Open is gated and has the widest attached policy. The Hub page asks for your full legal name, date of birth and organisation before download. The licence attaches a Usage Policy as Exhibit A, and you must follow it. The "Built with Muse Spark Open" notice applies to products that use the model, with no size threshold. The licence also defines a Derivative as "any AI model created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software". A distilled student model is therefore covered. The Hub card adds that this includes models trained on its outputs, which is worth confirming with counsel.

Kimi and Qwen scale their naming rule, and Muse and Mistral do not. Kimi and Qwen apply the naming rule only to very large products. Muse applies it to any product. Mistral has no naming clause in the text we read.

Which open-weight licences restrict hosting for others?

Four of the five families do: Mistral Large 4, Muse Spark Open, Kimi K3.1 Pro and both custom Qwen licences. The test is whether you give outside parties access to the model itself.

The Mistral and Kimi files define Model as a Service in nearly the same words. Mistral's text covers "giving a third party access to the Model or its derivative works for inference or post-training by any means". Kimi's covers giving a third party access to "language model inference or fine-tuning (e.g., via API)". Both add a control test: the third party must be able to exercise "meaningful control over the inputs, parameters, or training data."

Both also name two things that are not Model as a Service. These are end-user products with model capabilities "solely embedded within specific features or harnesses", and "mere relaying of requests to models hosted by others". The Qwen files describe it much the same way and carry the relaying exclusion. Meta words it differently. Its licence defines the activity as providing a third party with "access to artificial intelligence models as a product or service, via API or a similar interface".

Do not apply one licence's embedded-feature exemption to another. Kimi expressly excludes certain embedded features. Qwen does not include that exclusion and separately restricts AI Work Assistant businesses. An API that lets customers send their own prompts and set their own parameters is the clearest case of hosting. Anything in between is a question for counsel.

The consequence differs by licence:

• Mistral: any Model-as-a-Service business needs a separate commercial licence. The text says a licensee "must enter into a separate commercial license with Mistral". We found no revenue floor.
• Kimi: only above US$20M in revenue over any 12 consecutive months, counted across the licensee and its affiliates.
• Muse: only above US$20M in gross revenue over any consecutive 12 months, again counting affiliates.
• Qwen Community License 1.0: any Model-as-a-Service or AI Work Assistant business, with no threshold in the text. Qwen3.8-Max adds a US$50M threshold.

The Qwen "AI Work Assistant" limb catches something the others do not. The licence defines it as a standalone AI product "primarily designed for AI-assisted coding or office productivity". If you plan a coding assistant on Qwen3.8-Flash-Next, read that definition before you build.

Several files also exempt use through the licensor's own channels. Mistral exempts "any use of the Model accessed through Mistral's official products". Kimi exempts its own official products and "certified inference partners". If you buy the model as an API from the lab, these hosting conditions are not your problem.

Is Mistral Large 4 open source?

Mistral Large 4 is open-weight under a licence Mistral calls "a modified MIT license". It is not the plain MIT licence, and we would not call it open source.

Mistral's announcement says the model is "free to use, modify, fine-tune, and commercialize, with a single condition". The condition is the one above: businesses that operate a Model-as-a-Service offering need a separate commercial licence. Mistral's page tells self-hosters to download the weights and deploy them "with vLLM or SGLang", and describes the licence as a modified MIT licence.

The licence file is titled the "Mistral Ecosystem Open-Weight License". It is short, about 2.6 kB on the Hub, with an AS-IS warranty disclaimer and, in the text we read, no termination or trademark clause.

If you built on an earlier Mistral model under different terms, do not assume Large 4 carries the same ones.

Mistral's page calls it a "1-trillion-parameter" model but gives no active-parameter count, no context length and no date, so we rely on the licence text alone.

Can you run these models as a service inside your own walls?

Under the licences we read, internal use is exempt from the hosting conditions in every custom file. The catch is how each file defines internal use.

Kimi's file defines it as "any use that does not make the Software, its outputs, or its underlying capabilities available to third parties". Qwen's text says the same, including outputs. Mistral's definition refers to use that does not make "the Model or its derivative works available to third parties", in the clause summary we read. Muse's licence exempts "internal use, research, or any commercial activity that is not Model as a Service".

For sovereign and regulated buyers, this gives a workable pattern, with caveats:

• An internal platform for your own staff: within the internal-use exemption on all four custom licences. Confirm how each defines "third party".
• A shared service for several group companies: the licences count affiliates for revenue, and the definition of "third party" may exclude or include affiliates. Ask counsel.
• A hosted model for customers: this is Model as a Service on the wording above. In-region hosting by a cloud or integrator for its customers is hosting.
• Outputs shown to customers: the Kimi and Qwen wording on "outputs" means an internal tool whose answers reach customers needs a legal reading, not an assumption.

If this question blocks you, Apache 2.0 removes it. Beam and Qwen3.8-27B carry no hosting condition in the text we read. That is the main reason Beam matters to sovereign buyers, which we cover in our Beam post, alongside the hardware needed to run it. For the largest releases, our GPU requirements post covers what each node costs to stand up.

What should legal and procurement check?

Treat each model ID as an asset with a licence attached. This checklist works for any open-weight model, not only these five.

1. Open the file, not the model card. The card often names the licence without reproducing it. For Qwen3.8-27B the page shows only a tag, and we had to open the repository file.
2. Record the licence version and the date you read it. Hugging Face files can change. Store the licence next to the weight hashes.
3. Classify your use: internal, embedded in your product, or access for third parties. The definitions above decide which.
4. Check for a revenue threshold. Compare it with your group's revenue, not one entity's, because Kimi, Muse and Qwen count affiliates.
5. Check for a naming rule. If attribution applies, assign responsibility for the notice and check the permitted locations.
6. Read the usage policy. Muse attaches one. It affects what your users may do with the product.
7. Look at derivatives. If you fine-tune or distil, ask whether your resulting model inherits the licence terms. Muse defines Derivatives broadly.
8. Check the contracting entity. The Muse card says the Meta entity depends on your region.
9. Plan the separate agreement. If you need one, start the conversation before you build. We found no published price for any of the separate agreements.
10. Keep a swap path. Put model IDs behind aliases, so a licence change is a configuration change.

A last warning on sources. Press summaries of licences are loose, and a model card's one-line label is looser still. The files spell the conditions out, and the details matter.

Where Swfte sits

Licence review is a human task, and Swfte does not provide legal advice. What a platform can do is keep model choice flexible, so a licence finding is not a rewrite.

Swfte Connect offers one API across many providers, with smart routing. You can bring your own keys from any provider, or use Swfte's pooled access. When a provider is out of funds, rate-limited or down, Connect can fall back to another provider in the chain.

Note one point from the licences above. The Mistral, Kimi and Qwen files describe "mere relaying of requests to models hosted by others" as outside Model as a Service. That is those licensors' wording, and it concerns the party that hosts the model. It is not a statement about any Swfte deployment, and each customer should read the licence for the model it routes to.

For buyers who want models inside their own boundary, the enterprise deployment generator produces Terraform, Helm and Docker Compose bundles. Generating one requires a licence key, and dedicated deployment is available on request. Models not in the Connect catalogue are treated as bring-your-own-key with an optional base URL, so a custom OpenAI-compatible endpoint, such as your own vLLM server, can sit behind routing, failover and usage caps. Swfte-hosted customer data is held in AWS eu-west-1 (Ireland).

The decision rule: if you must host a model for others, prefer an Apache-licensed one or budget for a separate agreement before you build.

---

Related: Reflection AI Beam under Apache 2.0, What a trillion-parameter open model needs, Model supply chain security: hashes and licences, Open weights vs proprietary AI models.

Sources: Hugging Face, Beam 1.0 LICENSE · Hugging Face, Mistral Large 4 LICENSE · Mistral AI, Mistral Large 4 · Hugging Face, Muse Spark Open LICENSE.md · Meta, Muse Spark Open (15 October 2026) · Hugging Face, Kimi K3.1 Pro LICENSE · Hugging Face, Qwen3.8-27B LICENSE · Hugging Face, Qwen3.8-Flash-Next LICENSE · Hugging Face, Qwen3.8-2.4T-A95B LICENSE · Apache Software Foundation, Apache License 2.0 · Open Source Initiative, Open Source Definition · Open Source Initiative, Open Source AI Definition 1.0, accessed 9 November 2026.

This is not legal advice. Licence files can change, so confirm the current text before you rely on it. Clause summaries for the Mistral file come from a fetched reading that returned short quotes, and we quote only those fragments. We did not verify OSI's view of any custom licence here, and we have not seen the terms of any separate commercial agreement.