Open LLM licences for commercial use: Apache 2.0, MIT, Llama, Qwen, Mistral

Model selected, licence clarified, operation still open? WZ-IT runs open-weight models on your own infrastructure and documents model version, quantization and licence status as part of the model approval. Models run on the AI Cube on your premises or on managed GPU servers, operated through LLM hosting. Book an introductory call · Explore LLM hosting

Open language models can be downloaded and run on your own hardware. Whether a company may also use a particular model commercially, adapt it or provide it to customers is not stated in the model name but in the licence file of the specific version. Terms can differ within a single model family, and some licences contain revenue thresholds, naming obligations or regional exclusions. This article classifies the common licence types, summarises the terms of important models and answers the questions that arise with quantization, fine-tuning and hosting for third parties. As of October 2026.

Note: This article is a technical assessment based on the published licence texts and model cards. It is not legal advice and does not replace a legal review of the individual case.

Table of contents

• Open weight is not the same as open source
• Licence types at a glance
• Licences of important models
• Llama: community licence with an EU exclusion
• Qwen, Mistral and GLM: one family, several licences
• NVIDIA licences and non-commercial models
• Practical questions: quantization, fine-tuning, hosting, outputs
• Licence review as part of model approval
• What this means for your project

Open weight is not the same as open source

"Open weight" only describes availability: the trained weights of a model can be downloaded publicly. Which rights come with them is set by the accompanying licence. It decides three questions that matter to companies:

The licence shown in the header of a Hugging Face model card is a good first indication, but not binding. If it says other, the model has its own licence whose text has to be read. What is binding is the licence file of the version in use, including the version from which a quantization or fine-tune was derived.

The licence governs usage rights to the model. Obligations under the AI Act, data protection law or professional secrecy apply independently. The regulatory framework is described in The EU AI Act for companies.

Licence types at a glance

Apache 2.0 is the most widely used permissive model licence. Section 4 requires, on distribution, a copy of the licence, change notices in modified files, retention of copyright and patent notices and, where applicable, carrying over the NOTICE file. Section 3 contains a patent licence that ends if you file patent litigation over the work. Section 6 expressly grants no trademark rights, so model names may not be used as your own product brand (Apache License 2.0).

MIT is even shorter: the copyright and permission notice must be included in all copies, with no further conditions. "Modified MIT", by contrast, means the vendor has added conditions, and these vary widely (see below).

Licences of important models

Licences of specific model versions, checked against licence file and model card, as of October 2026:

Which of these models fit which hardware is covered in Which LLM to self-host? and GPU and VRAM sizing for LLMs.

Llama: community licence with an EU exclusion

The Llama 4 Community License (effective 5 April 2025) permits use, modification and distribution, but attaches conditions:

For companies in the EU, one rule matters more than the user threshold. The Llama 4 Acceptable Use Policy does not grant the rights under section 1(a) for multimodal models included in Llama 4 to individuals domiciled in, or companies with their principal place of business in, the EU. End users of a product or service incorporating such a model are exempt. According to the model card, the Llama 4 models are natively multimodal. The same rule already appears in the Llama 3.2 usage policy for its multimodal models.

For a company based in Germany or elsewhere in the EU that wants to run Llama 4 itself, this is an exclusion that has to be assessed legally before deployment. Comparable models under Apache 2.0 or MIT avoid the question.

Qwen, Mistral and GLM: one family, several licences

With several vendors, the licence depends on the model size or the release. A blanket approval such as "Qwen is Apache 2.0" or "Mistral is Apache 2.0" is therefore not reliable.

Qwen. The mid-sized models Qwen3.8-27B and Qwen3.6 are under Apache 2.0. Qwen3.8-Flash-Next is under the Qwen Community License 1.0: if the licensee or an affiliate runs a model-as-a-service or AI work assistant business, it needs a separate licence before any commercial use. Internal use is exempt as long as neither the model nor its outputs or capabilities are made available to third parties. Model-as-a-service here means third-party access to inference or fine-tuning with meaningful control over inputs, parameters or training data. The Qwen3.8-Max License for the 2.4-trillion-parameter model applies the same requirement only above USD 50 million revenue in twelve months. Both licences additionally require the model name to be displayed prominently if the product has more than 100 million monthly active users or USD 20 million monthly revenue.

Mistral. Mistral Small 4, Mistral Large 3, Ministral 3 and Devstral Small 2 are under Apache 2.0. Mistral Medium 3.5 and Devstral 2 (123B) are under a Modified MIT License with a hard limit: if the company's global consolidated monthly revenue exceeded USD 20 million in the preceding month, no rights under the licence may be exercised. This applies to the model and all derivatives, including internal use. Older models such as Codestral 22B v0.1 are under the Mistral AI Non-Production License and are not cleared for production use.

GLM. GLM-5.3-Flash and GLM-5.2 are under MIT. The larger GLM-5.3 has its own licence: if the licensee runs a model-as-a-service business and revenue exceeds USD 10 billion over twelve months, a security review by Z.AI is required before commercial use. For mid-sized companies this has practically no effect, but it changes the licence category.

NVIDIA licences and non-commercial models

NVIDIA uses several model licences that differ in one point:

The older NVIDIA Open Model License terminates the rights automatically if technical limitations or safety guardrails of the model are bypassed or weakened without a substantially similar replacement. This matters for fine-tuning that changes safety behaviour. All three NVIDIA licences claim no ownership of model outputs and terminate if you file patent or copyright litigation over the work.

Non-commercial licences rule out company use without a separate agreement. This mainly affects auxiliary models in RAG systems: jina-embeddings-v3 and jina-reranker-v2-base-multilingual are under CC BY-NC 4.0. Embedding and reranker models are therefore checked just like the language model. Permissively licensed alternatives are listed in Hybrid search and reranking.

Practical questions: quantization, fine-tuning, hosting, outputs

Quantization. A model in FP8, AWQ, GPTQ or GGUF is a modified version of the weights. The licence stays the same; quantization does not remove any conditions. For community quantizations on Hugging Face, the base model is therefore checked, not only the information on the quantizer's page. The technical formats are explained in LLM quantization.

Fine-tuning. A fine-tune, for example a LoRA adapter, is a derivative work of the base model. As long as it stays within the company, Apache 2.0 and MIT add no further obligations. If it is published or passed on to customers, the distribution obligations apply, and for Llama also the naming rule.

Hosting for third parties. Apache 2.0 and MIT tie their obligations to the distribution of copies. Providing a model only via an API does not distribute any weights. Vendor licences, by contrast, sometimes address this case explicitly: Llama requires "Built with Llama" for services made available as well, and the Qwen Community License and the GLM-5.3 License address model-as-a-service directly.

Outputs. None of the licences reviewed claims rights to the generated text. Google states in the Gemma Terms of Use that it claims no rights in outputs, as does NVIDIA in its model licences. Restrictions concern the use of outputs to train other models. Whether copyright arises in the outputs themselves is not a question of the model licence.

Licence review as part of model approval

A licence review is not a one-off step but part of the model approval. A fixed scheme works well:

1. Pin the specific version: repository, model ID and revision (commit hash) of the base model, not just the family name.
2. Read the licence file: LICENSE and, where present, NOTICE, USAGE_POLICY or an Acceptable Use Policy. The entry in the model card header is not sufficient.
3. Check the derivation chain: for quantizations, fine-tunes and distillations, identify the original model and apply its licence.
4. Assign the deployment scenario: internal, customer product, service for third parties, distribution of weights. Check the obligations from the table above for each scenario.
5. Check thresholds: revenue and user thresholds (Mistral Modified MIT, Qwen, Kimi, Llama) against your own figures, including affiliates.
6. Include auxiliary models: treat embedding, reranker, vision and speech recognition models the same way.
7. Document the result: record licence status with date, reviewer and deployment scenario in the model register; review again when the model changes.

The same model register serves as the basis for updates and rollbacks in operation, because it records which version was approved with which licence status.

What this means for your project

For most company use cases, capable models are available under Apache 2.0 or MIT, for example gpt-oss, Gemma 4, Qwen3.8-27B, Mistral Small 4, DeepSeek V4 or Granite 4.2. With these licences, internal use, quantization and fine-tuning remain free of additional obligations. Vendor licences with revenue thresholds, naming rules or regional exclusions, on the other hand, require a legal assessment before deployment.

WZ-IT runs open-weight models with vLLM or Ollama on the AI Cube or on managed GPU servers with NVIDIA RTX PRO 4000 Blackwell (24 GB) or RTX PRO 6000 Blackwell Max-Q (96 GB). In managed AI operations, model version, quantization and licence status are documented per model and carried forward with updates. The legal assessment remains the responsibility of the company and its legal counsel. Support, consulting and implementation by WZ-IT.

Which models suit which task is described in Which LLM to self-host?, and a comparison of current model families is provided in Llama 4 vs. Qwen 3.5 vs. DeepSeek V4. Anyone who prefers to use models via European providers rather than running them in-house will find an overview in European LLM APIs compared.

Sources

• Apache License, Version 2.0
• Meta, Llama 4 Community License Agreement
• Meta, Llama 4 Acceptable Use Policy
• Meta, Llama 4 Model Card
• Meta, Llama 3.2 Acceptable Use Policy
• Google, Gemma 4 License
• Google, Gemma 4 announcement (2 April 2026)
• Google, Gemma Terms of Use
• Qwen3.8-27B, model card
• Qwen3.8-Flash-Next, Qwen Community License 1.0
• Qwen3.8-2.4T-A95B, Qwen3.8-Max License
• Qwen2.5-72B-Instruct, Qwen License
• Qwen2.5-3B-Instruct, Qwen Research License
• OpenAI, gpt-oss-120b model card
• Mistral Small 4, model card
• Mistral Medium 3.5, Modified MIT License
• Devstral 2 123B, Modified MIT License
• Mistral AI Non-Production License
• DeepSeek V4 Pro, MIT License
• Z.AI, GLM-5.3 License
• Z.AI, GLM-5.3-Flash MIT License
• Moonshot AI, Kimi K2.6 Modified MIT License
• Xiaomi, MiMo-V2.6 model card
• IBM, Granite 4.2 model card
• Swiss AI, Apertus model card
• NVIDIA Nemotron Open Model License
• NVIDIA Open Model Agreement
• NVIDIA Open Model License
• OpenMDW License 1.1
• Jina AI, jina-embeddings-v3 model card
• Jina AI, jina-reranker-v2-base-multilingual model card