In this guide · 8 sections
- 01Open source vs open weight: why the word matters
- 02Licence types at a glance (October 2026)
- 03Apache 2.0 and MIT: what you still have to do
- 04Custom community licences: the clauses to look for
- 05Fine-tuning, redistribution and hosted APIs
- 06Before you build on an open model: a checklist
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Usually yes for a small business, but check the licence of the exact model version. Apache 2.0 and MIT models (gpt-oss, Gemma 4, many Qwen and Mistral models, DeepSeek) allow commercial use with notices. Custom licences add conditions, such as Llama's 700 million user limit and 'Built with Llama' attribution.
Open source vs open weight: why the word matters
An 'open-weight' model means you can download the trained model file and run it yourself. That says nothing about what you may do with it. The rules come from the licence attached to that specific release on Hugging Face or the vendor's site.
Some models use standard open source licences approved by the Open Source Initiative, such as Apache 2.0 or MIT. Others use custom 'community' licences written by the vendor. These can still allow commercial use, but they add conditions that a standard licence would not, and they are not on the OSI's list of approved licences.
For a small business, the practical question is simple: does the licence let you use the model in your product or internal tools, and what do you have to do in return?
Licence types at a glance (October 2026)
The table below groups common model families by licence type, from the licence fields and files on each model's Hugging Face page, checked on 19 October 2026. Vendors change licences between versions, so treat this as a starting point.
| Model (example version) | Licence | Commercial use | Main conditions |
|---|---|---|---|
| OpenAI gpt-oss-20b / 120b | Apache 2.0 | Yes | Keep licence and notices; state changes you make |
| Google Gemma 4 | Apache 2.0 | Yes | Same as Apache 2.0. Older Gemma 3 used the Gemma Terms of Use |
| Qwen3.8-27B | Apache 2.0 | Yes | Same as Apache 2.0 |
| Qwen3.8-Flash-Next | Qwen Community License 1.0 | Yes, with limits | Name display above 100M MAU or US$20M monthly revenue; separate licence for model-as-a-service or AI work-assistant businesses |
| Mistral Small 4, Mistral Large 3 | Apache 2.0 | Yes | Same as Apache 2.0 |
| Mistral Medium 3.5 | Modified MIT | Yes, below a revenue cap | No rights if your company's monthly revenue exceeds US$20 million |
| DeepSeek V4 | MIT | Yes | Keep the copyright and licence notice |
| Llama 4 Scout / Maverick | Llama 4 Community License | Yes, with limits | 'Built with Llama' attribution, acceptable use policy, separate licence above 700M MAU |
| SEA-LION v4.8 (Nemotron-based) | MIT | Yes | Keep notices; check the base model's terms too |
Apache 2.0 and MIT: what you still have to do
Apache 2.0 and MIT are the easy cases. Both allow commercial use, changes and redistribution. Apache 2.0 also includes an explicit patent licence from contributors.
- Keep the licence text and copyright notices when you share the model or a fine-tuned version
- Under Apache 2.0, mark files you changed and pass on any NOTICE file
- You do not need to publish your own code or fine-tuned weights
- Neither licence gives you trademark rights, so do not imply the vendor endorses your product
- Running the model internally, without sharing the weights, triggers almost none of these duties
The licence covers the model only. Your use of the outputs still has to follow normal law, including the PDPA if you feed it personal data, and copyright if outputs copy someone else's work.
Custom community licences: the clauses to look for
Custom licences tend to reuse the same handful of clauses. Once you know them, a licence file takes ten minutes to check.
User or revenue thresholds
The Llama 4 Community License says that if your products had more than 700 million monthly active users in the month before the Llama 4 release date, you must ask Meta for a licence. Mistral's Medium 3.5 licence removes all rights if your company's monthly revenue exceeded US$20 million in the preceding month. Neither threshold affects a typical Singapore SME today.
Attribution and naming
Llama 4 requires you to display 'Built with Llama' on a related website, interface or documentation when you distribute the model or a product containing it. If you fine-tune and distribute a model based on Llama, its name must start with 'Llama'.
Business-model restrictions
The Qwen Community License 1.0 requires a separate licence from Qwen before you use the model commercially if you run a 'Model as a Service' business (giving third parties meaningful control over inference or fine-tuning via an API) or an 'AI Work Assistant' (an independent product mainly for AI coding or office productivity). Internal use that does not expose the model to third parties is excluded.
Acceptable use policies
Llama and older Gemma releases link to a separate acceptable use or prohibited use policy, which becomes part of the licence. Read it, because it can change after you download the model.
Fine-tuning, redistribution and hosted APIs
How you use the model changes which clauses apply. Most small businesses fall into the first two rows.
| How you use the model | What usually applies |
|---|---|
| Run it internally (staff chatbot, document search) | Very little: follow the acceptable use policy and keep notices |
| Use it inside a customer-facing app or website chat | Attribution clauses (for example 'Built with Llama'); business-model limits for some Qwen models |
| Fine-tune it and share the new weights | Licence copy, notices, naming rules (Llama), and the base model's licence carries over |
| Offer it to others as an API | Check for model-as-a-service clauses; standard Apache 2.0 and MIT allow it |
| Use it through a cloud provider's hosted version | The provider's terms apply on top of the model licence |
Fine-tuned and merged models inherit their base model's licence. A model on Hugging Face labelled MIT may be built on a Llama or Gemma base with stricter terms, so check the base model named on its card. Our guide on running a local LLM with Ollama shows where to find the licence for models you pull.
Before you build on an open model: a checklist
Licence risk is only part of the picture. Our guide on DeepSeek, Qwen and open model safety covers data and security questions, and SEA-LION for business looks at the Singapore-built option.
Open model licence check
Do it yourself or get help?
Checking a licence for an internal tool is a DIY job: find the model card, open the licence file, run the checklist above and keep a dated copy. For a customer-facing product, or anything you plan to sell or host for others, have a lawyer read the licence once and record their view.
If you want a custom AI tool built on an open model, our Custom Tool or Dashboard service (from S$1,990) includes a written scope that names the model, its licence and the attribution needed, with the system hosted in your own accounts and the source code handed to you.
Frequently asked questions
Yes, in most cases. The Llama 4 Community License allows commercial use, but you must follow Meta's acceptable use policy, display Built with Llama when you distribute the model or a product containing it, and request a separate licence if your products had over 700 million monthly active users. Check the licence for your exact version.
It depends on the model. Many Qwen models, such as Qwen3.8-27B, are released under Apache 2.0 and are free for commercial use. Others use the Qwen Community License 1.0, which requires a separate licence for model-as-a-service and AI work-assistant businesses and requires the model name to be shown on products above 100 million monthly users or US$20 million monthly revenue.
Many are. Mistral Small 4 and Mistral Large 3 are listed under Apache 2.0 on Hugging Face. Mistral Medium 3.5 uses a modified MIT licence that removes rights for companies with more than US$20 million in monthly revenue. Always check the licence on the specific model card you plan to use.
Only if the licence says so. Apache 2.0 and MIT require keeping notices when you distribute the model, not a credit on your website. Llama 4 requires Built with Llama to be shown on a related website, interface or documentation. Qwen's community licence requires the model name on screen only above 100 million monthly users or US$20 million monthly revenue.
Model licences mainly cover the weights and how you use them. Some, like Llama, also mention outputs when used to train other models. Your outputs still have to respect copyright, the PDPA and consumer law, so review generated content before publishing it, especially marketing claims and personal data.