Open-weight model licenses: what you can build with Gemma, Qwen, Muse, Llama and more

Open-weight model licenses decide what you may do with a model you can download: sell a product built on it, fine-tune it, offer it through an API or ship a renamed copy. "Open weights" only means the parameters are published; the license can still add conditions. On October 23, 2026 the popular open-weight families range from plain Apache 2.0 or MIT, with almost no strings attached, to custom licenses with revenue thresholds, attribution rules and usage policies. This guide compares the licenses of Gemma 4, gpt-oss, Qwen3.8, DeepSeek V4, Meta's Muse and Llama 4 models and Mistral's open models, based on the license texts and model cards read on October 23, 2026.

The licenses at a glance

Table: open-weight model licenses (license texts and model cards, checked October 23, 2026)

Licenses change between versions of the same family: always read the license that ships with the exact model you download.

Apache 2.0 and MIT: the simple cases

Apache 2.0 lets you use, modify and distribute the model, commercially or not. When you redistribute it, you must give recipients a copy of the license, mark the files you changed, keep the existing copyright and patent notices and pass on any NOTICE file. It also includes a patent license from contributors, which ends for you if you sue claiming the work infringes a patent. MIT is shorter: you may do almost anything as long as the copyright notice and the permission notice travel with copies.

Gemma 4 belongs in this group: Google's model card lists it under Apache 2.0. If you only remember one thing, it is this: an Apache 2.0 or MIT model can go into a paid product without asking anyone, which is what most businesses need.

Custom licenses: the clauses that matter

Qwen. Within one family the license varies by model. Qwen3.8-27B is Apache 2.0. Qwen3.8-Flash-Next uses the Qwen Community License 1.0, and Qwen3.8-2.4T-A95B the Qwen3.8-Max License. Both allow commercial use, require the model name to be shown prominently in products with more than 100 million monthly active users or 20 million dollars in monthly revenue, and require a separate license from Qwen to run a Model as a Service business or an independent coding or office productivity assistant. Under the Community License that applies at any size; under the Max License only when revenue passes 50 million dollars in twelve months. Internal use that does not expose the model to third parties is excluded.

Meta Muse Spark Open. Released on October 15, 2026, it allows use, fine-tuning, distillation and resale, with three conditions: include the license and its usage policy, display "Built with Muse Spark Open" in the product or its documentation, and, if you sell model access through an API and your group's gross revenue passes 20 million dollars in any twelve months, notify Meta and sign a separate agreement first. Meta claims no ownership of outputs. The usage policy lists prohibited uses and says the model is not intended for people under 18. With 816 billion parameters, it is not a laptop model: Meta says it runs on two Mac Studios with M5 Ultra. Our guide to running an LLM on a Mac covers smaller options.

Llama 4. The Llama 4 Community License is royalty-free, but companies whose products had more than 700 million monthly active users on the release date need a license from Meta. You must show "Built with Llama", include a notice file, and start the name of any model you train with Llama or its outputs with "Llama". Your license ends if you sue Meta claiming Llama infringes your rights, and California law governs the agreement.

Llama 4 in Europe

The Llama 4 acceptable use policy adds a clause European readers should not miss: for the multimodal models in Llama 4, the license rights are not granted to individuals living in the European Union or companies with their principal place of business there. End users of a product that includes those models are not affected. A Spanish company building on Llama 4's multimodal models therefore has a problem; one using Apache 2.0 or MIT models, or Muse Spark Open, does not face that clause. Our overview of Regulation (EU) 2024/1689 covers the separate obligations that EU law sets.

A checklist before you build

1. Read the model card's license field. On Hugging Face, "apache-2.0" or "mit" means a standard license; "other" means a custom one you must read.
2. Check thresholds. Monthly users, monthly revenue and annual revenue clauses decide whether you need a separate deal.
3. Plan the attribution. "Built with" lines and naming rules affect your interface and your model names.
4. Read the usage policy. It becomes part of the license where the license says so, as with Llama 4 and Muse Spark Open.
5. Check derivatives. Fine-tuning and distillation are usually allowed, but the conditions follow the new model.
6. Save a dated copy of the license you accepted, because new versions bring new terms.

If you run models yourself, Ollama vs LM Studio explains the software licenses of the tools as well. This page is our reading of the license texts, not legal advice; for a product that matters, have a lawyer read the license you rely on.