Open LLM licences for commercial use: Apache 2.0, Llama, Gemma, Qwen and others compared

Open LLM licences for commercial use compared

Open-weight models come under three kinds of terms. Permissive licences, Apache 2.0 and MIT, grant commercial use and ask mainly for the licence text and notices when you pass the weights on; gpt-oss-120b, Qwen3.8-27B, Gemma 4, Mistral Small 4 and DeepSeek V4-Flash are published this way. Custom licences from Meta, Google, NVIDIA, Qwen and Mistral grant broad use rights too but add use policies, naming rules, thresholds and, for Llama 4, a clause on the EU. A third set, such as releases under CC BY-NC 4.0, grants rights for non-commercial purposes only.

Terms are set per checkpoint: Qwen3.8-27B and Qwen3.8-Flash-Next have different licences, as do Gemma 4 and Gemma 3.

Model cards and licence files on Hugging Face, Meta’s and Google’s licence pages and NVIDIA’s agreement page, read on 10 October 2026.

Server sizing for these models is in our LLM hardware requirements by model, and the licences of EuroLLM, Bielik and PLLuM in our comparison of open LLMs for German, Polish and Romanian.

Apache 2.0 and MIT: what permissive licences ask for

Section 2 of the Apache License 2.0 grants a copyright licence that is “perpetual, worldwide, non-exclusive, no-charge” to reproduce, prepare derivative works of, sublicense and distribute the work. Its conditions sit in Section 4, Redistribution: recipients get “a copy of this License”, modified files carry “prominent notices stating that You changed the files”, and existing notices, including a NOTICE file, travel with the work. Section 3 ends the patent licence of anyone who starts patent litigation claiming that the work infringes, and Section 6 grants no right to the licensor’s trademarks.

The MIT licence has one condition: “The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.” Both texts disclaim any warranty and name no governing law. The Open Source Initiative lists both as approved licences, under the SPDX identifiers Apache-2.0 and MIT.

Mistral’s card for Small 4 calls Apache 2.0 an “Open-source license for both commercial and non-commercial use” and adds that the model must not be used in a way that infringes third parties’ rights. The gpt-oss-120b repository adds a usage policy: “By using OpenAI gpt-oss-120b, you agree to comply with all applicable law.” DeepSeek’s card for V4-Flash states: “This repository and the model weights are licensed under the MIT License.” The Gemma 4 cards link to Google’s copy of the Apache text.

Llama licences: Built with Llama, model names and 700 million users

Llama 3.3 and Llama 4 come under Meta’s community licence agreements, the Llama 4 one with a “Llama 4 Version Effective Date: April 5, 2025”. Meta’s model card calls it “A custom commercial license”. Section 1.a grants a non-exclusive, worldwide, non-transferable and limited licence to use, reproduce, distribute, copy, modify and create derivative works of the Llama Materials.

Under Section 1.b.i, whoever distributes or makes available the materials, a derivative or a product or service containing them must “provide a copy of this Agreement with any such Llama Materials” and display “Built with Llama” prominently on a related website, user interface or documentation. A model created, trained or fine-tuned with the materials or their outputs that is distributed or made available must carry “Llama” at the beginning of its name. Section 1.b.iv makes Meta’s Acceptable Use Policy part of the agreement, and Section 7 applies California law.

Section 2 applies if the products or services of the licensee or its affiliates had more than 700 million monthly active users in the calendar month before the release date: “you must request a license from Meta, which Meta may grant to you in its sole discretion”. The count covers the products or services made available by or for the licensee or its affiliates. The Llama 3.3 agreement of 6 December 2024 has the same attribution, naming and user clauses.

The Llama 4 clause on multimodal models in the EU

Meta’s model card describes the Llama 4 collection as “natively multimodal AI models”, and it lists text and image as the input of Scout. The Llama 4 Acceptable Use Policy, which shows no date on its page, states: “With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union.” The next sentence reads: “This restriction does not apply to end users of a product or service that incorporates any such multimodal models.”

The Llama 3.3 agreement contains no comparable territorial clause. Our Llama 4 hardware guide covers memory and GPUs for both models.

Model and GPU selection is part of the technical assessment of our Private AI/ML service, and its price is fixed before work begins. Tell us which models your departments want to run and for how many users.

Gemma Terms of Use for Gemma 3 and earlier models

Gemma 3, Gemma 3n, EmbeddingGemma and the other models in the appendix stay under the Gemma Terms of Use, last modified on 1 April 2026. The text names no user or revenue threshold. Section 3.2 bars use for the restricted uses of the Gemma Prohibited Use Policy and “in violation of applicable laws and regulations”, and Google “reserves the right to restrict” usage it reasonably believes violates the agreement.

The Prohibited Use Policy, last modified on 21 February 2024, lists “Making automated decisions in domains that affect material or individual rights or well-being”, with finance, legal, employment, healthcare, housing, insurance and social welfare among its examples.

Model Derivatives include, besides modifications, any model created by transfer of patterns of Gemma’s weights, parameters, operations or Output “in order to cause that model to perform similarly to Gemma”, with distillation and synthetic data as examples. Whoever reproduces or distributes Gemma or a derivative passes the use restrictions on as an enforceable provision, gives recipients a copy of the terms, marks modified files and ships a Notice file. Our Gemma hardware guide sizes the models.

NVIDIA, Qwen and Mistral licences with their own conditions

The NVIDIA Open Model License Agreement, last modified on 24 October 2025, governs NVIDIA models such as gpt-oss-puzzle-88B, which its card describes as “derived from OpenAI’s gpt-oss-120b”. NVIDIA states in it that “Models are commercially usable”, claims no ownership of outputs and allows derivative models to be created and distributed. Under Section 3.1, whoever distributes the model passes on a copy of the agreement and a Notice file reading “Licensed by NVIDIA Corporation under the NVIDIA Open Model License”. Under Section 2.1, rights end automatically for anyone who bypasses or disables “any technical limitation, safety guardrail” in the model without a substantially similar guardrail. Rights also end on patent or copyright litigation over the model, and US and Delaware law applies.

Qwen3.8-Flash-Next comes under the Qwen Community License 1.0. Its clause 2 states: “If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.” Model as a Service “means giving a third party access to language model inference or fine-tuning” in a way that lets the third party “exercise meaningful control over the inputs, parameters, or training data”. An AI Work Assistant is an independent product “primarily designed for AI-assisted coding or office productivity”, and the definition excludes single-purpose tools and assistants built into products with another main purpose. Internal use that makes neither the software, its outputs nor its model capabilities available to a third party is exempt. Clause 1 asks products with more than 100,000,000 monthly active users, or above a monthly revenue threshold, to display the model name prominently.

Mistral Medium 3.5 comes under a “Modified MIT License”. Its second section begins: “You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company” (or that of your employer) exceeds an amount the licence states for the preceding month. The restriction “applies to the Model and any derivatives, modifications, or combined works based on it”, whether Mistral AI or a third party provides them.

Non-commercial and research releases

CC BY-NC 4.0 grants its rights “for NonCommercial purposes only” and defines NonCommercial as “not primarily intended for or directed towards commercial advantage or monetary compensation.” Teuken-7B-instruct-v0.6 is published under it, and its card states that the model “is designed for private, non-commercial, research, and educational use”.

Where each licence answers the common questions

This table maps the questions of a legal or procurement review to the clause that answers them.

Apache License 2.0 (apache.org), MIT licence (opensource.org), Llama 4 Community License Agreement and Acceptable Use Policy, Gemma Terms of Use, NVIDIA Open Model License Agreement, read on 10 October 2026.

Reviewing a model licence before a company-wide rollout

For an assistant used by 500 to 2,000 employees, the licence review belongs to the model decision, before the pilot.

1. Record the repository, revision, licence file and date read for every checkpoint you deploy, quantised versions from third parties included.
2. Read the licence of the checkpoint you run, not only that of its base: gpt-oss-120b is Apache 2.0, while NVIDIA’s derived gpt-oss-puzzle-88B is under the NVIDIA Open Model License.
3. Add the documents a licence incorporates by reference, such as Meta’s Acceptable Use Policy or Google’s Prohibited Use Policy, with the date each was read.
4. Apply thresholds to the whole company with its affiliates, since the Llama and Qwen licences count the licensee and its affiliates.
5. List whether weights, fine-tunes or services built on them will leave the company or be made available to others, for example subsidiaries or customers, because the notice, naming and flow-down clauses attach to distribution and, in the Llama licences, to making available.

The legal assessment of each licence against your use cases is for the company’s legal department. A company that builds an assistant on an open model and puts it into service under its own name, which includes supply for own use, can also be its provider under the AI Act, as our guide to the EU AI Act for companies using LLMs explains.

Our Private AI/ML service deploys open and commercial models on your servers or in an EU data centre, with a query log. Send us your model shortlist and the processes it will serve through the form below.

What we do

Our Private AI/ML service deploys open and commercial models on-premise or in a Tier-3 data centre in Lithuania, with vLLM, Ollama or NVIDIA AI Enterprise, and logs queries and answers so that security and legal see who accesses what. The first call is free of charge. The legal compliance assessment stays with your legal department, and we deliver the technical part. Engineering is by our partner Vixen.UNO, on one EU contract.

FAQ