Short answer: there is no single "open-weight licence". The licence is set per model, not per vendor, so you have to read it on each model card. As of 20 November 2026, Gemma 4 is under Apache 2.0, which is the cleanest case for commercial use. Llama 4 is open-weight under Meta's community licence, with a 700 million monthly active user threshold, a "Built with Llama" display duty and a naming rule. Qwen and Mistral are mixed: some models are Apache 2.0, others carry revenue caps, display duties, a ban on offering the model as a service, or a non-commercial limit. Many software companies will be able to use many of these models commercially, but the exceptions are exactly where deals and audits go wrong.
This post is a fact table plus a method. It summarises public licence terms we read on 20 November 2026, and it is not legal advice. Licences get revised, a vendor can license sibling models differently, and your own facts (group revenue, resale, EU users) decide the answer. Have counsel read the licence text on the model card before you ship.
What does open-weight mean, and is it the same as open source?
An open-weight model is one whose trained weights you can download and run on your own servers. That is a statement about access. It says nothing about what you may do with the weights. The licence decides that.
Some licences are permissive. Apache 2.0 is a standard permissive licence with no user or revenue cap. Others are custom community licences written by the model vendor. They let most people use the model commercially but add conditions: size thresholds, attribution, naming rules, an acceptable use policy, or limits on reselling the model as a service. Do not call a model "open source" in a contract, a questionnaire or a board paper unless its licence is an open source licence. Llama, for example, is best described as open-weight under Meta's community licence.
This matters most if you are building a private deployment, which is what our Own Your AI service helps with. The licence is the first filter before you spend time on benchmarks, GPUs or fine-tuning.
Which licence does each model family use?

The table below is the one to put in front of procurement. Each row is a model or a small group of models, because vendors license siblings differently. Read it as of 20 November 2026 and check the model card for the version you download.
| Model | Licence | Main commercial conditions |
|---|---|---|
| Gemma 4 | Licence: Apache 2.0 | Main commercial conditions: No user cap or revenue cap in the licence |
| Earlier Gemma | Licence: Gemma Terms of Use (1 April 2026) | Main commercial conditions: No user cap. Restrictions must pass downstream. Google may restrict use "remotely or otherwise" |
| Llama 4 (Scout, Maverick) | Licence: Llama 4 Community License (effective 5 April 2025) | Main commercial conditions: Licence request to Meta above 700M monthly active users. "Built with Llama" display. Derivative names start with "Llama". Acceptable Use Policy applies |
| Qwen3.8-27B | Licence: Apache 2.0 | Main commercial conditions: No user cap or revenue cap in the licence |
| Qwen3.8-Flash-Next | Licence: Qwen Community License 1.0 | Main commercial conditions: Display the model name in the UI above 100M MAU or US$20M monthly revenue. Model-as-a-Service or "AI Work Assistant" business needs a separate licence. Internal use exempt |
| Qwen3.8-2.4T-A95B | Licence: Qwen Max License | Main commercial conditions: Separate licence needed only above US$50M revenue over 12 months |
| Qwen-Image-2.1 | Licence: Research License | Main commercial conditions: Non-commercial only |
| Mistral Small 4, Leanstral, Shieldstral | Licence: Apache 2.0 | Main commercial conditions: No user cap or revenue cap in the licence |
| Mistral Medium 3.5 | Licence: Modified MIT | Main commercial conditions: Not usable if your company's global monthly revenue was above $20M in the previous month |
| Mistral Large 4 | Licence: Mistral Ecosystem Open-Weight License | Main commercial conditions: A Model-as-a-Service business needs a commercial licence. Internal use and features with the model embedded are exempt |
| Mistral-Large-2411, Codestral-22B | Licence: Mistral research or non-production licences (tagged mrl and mnpl) | Main commercial conditions: Research or non-production use. Read the text before any commercial use |
Two warnings. First, never write "Qwen is Apache 2.0" or "Mistral is open source" as a blanket statement. Both vendors publish models under several licences, and the same family can include a permissive model and a restricted one. Second, newer Meta models are not covered here. This post covers Llama 4 only, because that is the Meta release whose licence we read in full.
What are the Llama 4 licence conditions for commercial use?
Llama 4 (Scout and Maverick) is under the Llama 4 Community License, effective 5 April 2025. Four points matter for a commercial team.
- A user threshold. A company with more than 700 million monthly active users in the preceding calendar month must request a licence from Meta. Ask whether affiliates count towards it.
- Attribution. If you distribute the model or a product that includes it, the licence asks you to prominently display "Built with Llama".
- Naming. Distributed derivative models, such as a fine-tune you publish, must have a name that starts with "Llama".
- The Acceptable Use Policy. It applies to every use. For the multimodal models, the policy withholds the licence grant from individuals and companies based in the EU. End users of products built on those models are exempt.
The EU point is the one Indian SaaS teams miss. If you build a product on a multimodal Llama 4 model and sell it to EU customers, the end users are exempt, but a company based in the EU that uses the model directly is not. The policy ties this exclusion to the multimodal models, so check which model you are using. If you have EU customers or an EU entity, put this to counsel before choosing a multimodal Llama 4 model.
The licence text, not this summary, is what binds you.
Is Gemma free for commercial use?
It depends on the version, which is why "Gemma" alone is not an answer. Gemma 4 is under Apache 2.0, so the licence adds no user cap, revenue cap or UI display duty. Like any Apache 2.0 licence, it does ask you to pass on the licence text and notices when you redistribute the weights. That makes it the simplest model family on this list for a commercial product.
Earlier Gemma models are under the Gemma Terms of Use dated 1 April 2026. These terms have no user cap, but they require you to pass restrictions to downstream users, and they let Google restrict use "remotely or otherwise". That last phrase is the one a security or procurement reviewer will stop on. If your product depends on an earlier Gemma model, you cannot assume your right to run it is permanent and untouched.
The practical lesson: when a vendor moves a newer model to Apache 2.0, do not assume the older ones moved with it. Record the licence per model and per version.
Is Qwen Apache 2.0?
Some Qwen models are. Others are not. As of 20 November 2026:
- Qwen3.8-27B is Apache 2.0.
- Qwen3.8-Flash-Next is under the Qwen Community License 1.0. A product with more than 100 million monthly active users or more than US$20 million in monthly revenue must display the model name in its UI. A Model-as-a-Service business, or an "AI Work Assistant" business for coding or office productivity, needs a separate licence. Internal use is exempt.
- Qwen3.8-2.4T-A95B is under the Qwen Max License. A separate licence is needed only above US$50 million in revenue over 12 months.
- Qwen-Image-2.1 is under a Research License and is non-commercial only.
The line that catches teams is "AI Work Assistant". If you are building a coding assistant or an office productivity tool and you plan to put Qwen3.8-Flash-Next behind it, that is a use the licence singles out. Do not rely on the fact that your own revenue is small. Either choose an Apache 2.0 Qwen model or ask counsel whether you need the separate licence, and get it in writing if you do.
Is Mistral open source?

Mistral publishes models under several licences, so treat each model on its own terms.
- Mistral Small 4, Leanstral and Shieldstral are Apache 2.0.
- Mistral Medium 3.5 is under a Modified MIT licence. You may not use it if your company's global monthly revenue was above $20 million in the previous month. "Global" and "company" are the words to ask counsel about, especially if you are part of a group.
- Mistral Large 4 is under the Mistral Ecosystem Open-Weight License. A business that offers the model as a service needs a commercial licence. Internal use and features with the model embedded are exempt.
- Mistral-Large-2411 and Codestral-22B carry licences tagged mrl and mnpl on Hugging Face. We have not read those texts in full, so treat them as research or non-production licences and do not build a commercial product on them without reading the text and getting counsel's view.
The revenue cap on Medium 3.5 is a monthly test against the previous month. A company that crosses the line can find that a model it was already running no longer qualifies. That is a reason to track the licence in a register and to have a fallback model, which we cover below.
What does Model-as-a-Service mean in these licences?
Qwen3.8-Flash-Next and Mistral Large 4 both limit offering the model itself as a service. In plain terms, the worry is a business whose product is access to the model: an API, a hosted chat service, or a hosted assistant that is mostly the model with a thin wrapper. Using the model inside your own product, or for your own staff, is treated differently. Qwen's Community License exempts internal use. Mistral Large 4's licence exempts internal use and features with the model embedded.
The licences give their own definitions, and those definitions are what count. Use this test to decide whether you need to ask counsel:
- Do customers pay mainly for access to the model's output, or for a workflow that happens to use it?
- Could a customer call your endpoint as a generic model API?
- Is your product a coding assistant or an office productivity assistant (the Qwen "AI Work Assistant" wording)?
- Do you resell capacity from a model you host to other companies?
A yes to any of these means read the licence's service clause closely before you launch.
How should a procurement or security team check an open-weight licence?
Put the same questions to every model, in the same order. Pass the answers to counsel for sign-off.
| Check | Why it matters | Where to find it |
|---|---|---|
| Exact licence name and version | Why it matters: Siblings in one family use different licences | Where to find it: The model card and the LICENSE file |
| User or revenue threshold | Why it matters: A cap can make the model unusable or trigger a separate licence | Where to find it: Licence text |
| Whose revenue counts | Why it matters: Group, company or product level changes the answer | Where to find it: Licence text, then counsel |
| Commercial use allowed at all | Why it matters: Some models are research or non-commercial only | Where to find it: Licence text, README |
| Model-as-a-Service or resale limits | Why it matters: Your product may fall inside the restricted category | Where to find it: Licence text |
| Attribution and naming duties | Why it matters: A UI display duty or a required name prefix is a product change | Where to find it: Licence text |
| Acceptable use policy | Why it matters: It is a separate document that also binds you | Where to find it: Vendor use-policy page |
| Geographic exclusions | Why it matters: Llama 4 multimodal withholds the grant from EU-based individuals and companies | Where to find it: Use-policy page |
| Downstream restrictions | Why it matters: Earlier Gemma terms require passing restrictions on | Where to find it: Licence text |
| Remote restriction rights | Why it matters: Earlier Gemma terms allow Google to restrict use "remotely or otherwise" | Where to find it: Licence text |
Keep the licence text you accepted. Save a dated copy of the LICENSE file and the model card for the version you deploy, next to the model's revision identifier. If a vendor changes terms later, you can show what applied when you downloaded the weights.
How do you make licence compliance part of the engineering workflow?
Treat each model like a third-party dependency with its own licence. A short register is enough. It is also what a security reviewer will ask for when a customer asks which models you run and where, a question that appears in AI security questionnaires.
- Model register. One row per model in production: name, version, source, licence name, a link to the saved licence copy, who approved it and when.
- Pinned versions. Deploy a specific revision, not "latest". The licence on the model card for that revision is the one you reviewed.
- Approval gate. A change to a new model or version goes through the same check as the table above before it reaches production.
- Threshold watch. If a licence has a revenue or user cap, put the cap and a review date in the register. Revisit it when your numbers or your group structure change.
- A fallback. Keep at least one Apache 2.0 model evaluated and ready, so a licence change or a threshold crossing does not become an outage.
- Customer-facing answers. Keep the one-line description of each model's licence ready for questionnaires and contracts. Say "open-weight under the vendor's community licence" where that is accurate, not "open source".
This register also feeds the rest of your governance. The AI governance work for selling to enterprises asks for the same inventory: which models, which versions, which data they touch. For choosing between candidate models on quality and cost, read our guide on evaluating open-weight models for enterprise use. For moving an existing API workload to an open-weight model, see how to migrate from an API to an open-weight model.
Does the licence change what you can do with Indian data?
Not directly. A model licence governs your rights over the weights. The Digital Personal Data Protection Act, 2023 governs the personal data you put through them. Running an open-weight model on infrastructure you control is one way to keep prompts out of a third-party processor, but the licence does not make you compliant and compliance does not depend on which licence the model has.
The two sets of obligations meet in one place: your records. A model register that names the licence, and a data map that names the data each model touches, give you one answer for an auditor or an enterprise buyer. If you want the India-hosting side, our Own Your AI in India page covers where these models can run.
Which open-weight model is the safest choice for a commercial product?
There is no single safe model, but there is a safe order of work. Start from the licence, not the leaderboard.
- Pick Apache 2.0 models first where quality is close enough. As of 20 November 2026 that includes Gemma 4, Qwen3.8-27B, and Mistral Small 4, Leanstral and Shieldstral. The licence adds no user cap, revenue cap or UI display duty (you still pass on the licence text and notices when you redistribute), so the review is short.
- Use community-licence models with open eyes. Llama 4, Qwen3.8-Flash-Next and Mistral Large 4 are usable commercially in many cases. The conditions above decide whether yours is one of them.
- Avoid research and non-commercial models for anything customer-facing. Qwen-Image-2.1 is non-commercial only, and the older Mistral models tagged mrl and mnpl are for research or non-production use.
- Re-check on every version change. The licence travels with the version, not the family name.
If you want a second pair of eyes on a model choice, a tech or architecture review can cover the licence register alongside hosting and data flows. Our tech audit service does this kind of review. Mono does not give legal opinions, so the final reading of any licence goes to your counsel.
Sources: Meta, Llama 4 Community License (dev.meta.ai/llama/llama4/license/) and Llama 4 Acceptable Use Policy (dev.meta.ai/llama/llama4/use-policy/). Google, Gemma 4 licence (ai.google.dev/gemma/docs/gemma_4_license) and Gemma Terms of Use dated 1 April 2026 (gemma.ai/resources/terms/). Qwen and Mistral model LICENSE and README files on Hugging Face (huggingface.co, per model organisation and model name). Licence summaries are not legal advice; read the licence on each model card. Last reviewed 20 November 2026.
FAQ
Can I use open-weight models commercially?
Often yes, but it depends on the model's licence. As of 20 November 2026, Gemma 4, Qwen3.8-27B and Mistral Small 4 are under Apache 2.0. Llama 4 and some Qwen and Mistral models add thresholds, display duties or service-resale limits. Qwen-Image-2.1 is non-commercial only. Read the licence on each model card.
Is Llama 4 open source?
No. Llama 4 is open-weight under Meta's Llama 4 Community License, not an open source licence. The licence requires a request to Meta above 700 million monthly active users, a "Built with Llama" display for distributors, and derivative model names that start with "Llama". The Acceptable Use Policy also applies.
Is Gemma 4 free for commercial use?
Gemma 4 is under Apache 2.0, which adds no user cap or revenue cap. Earlier Gemma models use the Gemma Terms of Use of 1 April 2026, which pass restrictions downstream and let Google restrict use "remotely or otherwise". Check which version you are running and keep a copy of its licence.
Are all Qwen and Mistral models Apache 2.0?
No. Qwen3.8-27B and Mistral Small 4 are Apache 2.0, but Qwen3.8-Flash-Next, Mistral Medium 3.5 and Mistral Large 4 carry custom conditions, and Qwen-Image-2.1 is non-commercial only. Both vendors license models individually, so record the licence per model and version.
What is the Model-as-a-Service restriction in Qwen and Mistral licences?
Some Qwen and Mistral licences require a separate commercial licence if your business is offering the model itself as a service. Qwen3.8-Flash-Next also names "AI Work Assistant" businesses for coding or office productivity. Internal use is exempt, and Mistral Large 4 also exempts features with the model embedded. Counsel should read the exact definitions.
Does a licence make a model compliant with DPDP?
No. A model licence sets your rights over the weights. The DPDP Act governs the personal data you process with them. Self-hosting an open-weight model can keep prompts inside infrastructure you control, but you still need notices, consent, safeguards and records. Legal sign-off on both goes to counsel.
