theLLMs

Last checked: 2026-05-25

Scope: Global. Licence terms checked on 2026-05-25; individual model licences can change between versions.

AI draft model: deepseek-v4-flash

AI review model: llm-editor (deepseek-v4-pro)

Key terms:

>_ large language model (LLM) >_ open weights >_ LoRA (Low-Rank Adaptation) >_ JSON mode >_ function calling / tool use >_ reasoning model

Meta Llama and open model licensing: what builders must check

Open-weight models like Meta’s Llama series are not “open source” in the traditional sense — and the difference matters for commercial use, redistribution, and liability.

Most teams discover this when they are already building on a model and a licence clause blocks their product path.

Quick answer

Open-weight model licences vary significantly. Meta’s Llama 2 used a custom licence that granted broad commercial use but restricted use by organisations with over 700 million monthly active users. Llama 3 moved to a permissive custom licence with fewer restrictions but still does not meet the Open Source Initiative’s definition of open source. Mistral’s models are released under Apache 2.0, which is a recognised open-source licence. Other models use custom licences, research-only licences, or Creative Commons variants.

The practical question is always: can you use this model in a commercial product, under what conditions, and what happens if the licence changes in a future version?

What “open” actually means for model weights

Traditional open-source software licences (Apache 2.0, MIT, GPL) grant specific rights: use, modification, redistribution, and sublicensing. Model licences often follow a different pattern.

Open weights means the model parameters are publicly downloadable. It does not mean the licence is open-source, that training data is disclosed, that you can fine-tune and redistribute, or that you are protected from patent claims.

The OSI’s Open Source AI Definition (2024) tried to clarify this, but most major model licences still do not comply.

What the major licences allow

Meta Llama 2 and 3 (custom licence): broad commercial use allowed, redistribution with attribution required, use by entities over 700M MAU requires Meta’s approval (Llama 2 — removed in Llama 3), acceptable use policy applies, no patent grant.

Apache 2.0 (Mistral 7B, Mixtral): broad commercial use, modification and redistribution allowed, patent grant included, no additional use restrictions, recognised open-source licence.

Custom research licences (various): use permitted for non-commercial research only, commercial use requires separate agreement, redistribution often prohibited.

Creative Commons (some smaller models): varies by variant — CC-BY allows commercial use with attribution, CC-NC prohibits commercial use.

Where teams get tripped up

  1. Assuming “open weights” equals “open source.” The terms matter for compliance. An Apache 2.0 model can be forked and redistributed without asking permission. A Llama model cannot.

  2. Missing the acceptable use policy. Meta’s Llama licence incorporates an Acceptable Use Policy that restricts certain applications — and that policy can change between model versions without direct notice.

  3. Building on a research-only licence with a commercial exit plan. Several popular small models are released under research-only terms. Using them in a product — even for internal prototyping linked to a commercial service — can breach the licence.

  4. Redistributing fine-tuned weights without checking the base licence. The base model’s licence applies to derived works. If the base model is under a research-only licence, your fine-tune is also research-only unless the licence explicitly allows derivative redistribution.

  5. Not tracking licence changes across model versions. A model series may switch licences between major versions (Llama 2 to Llama 3 changed the MAU clause). If you pinned a specific version, the old licence may still apply — but downstream users may not know which version’s terms they are under.

Practical decision check

Before building on an open-weight model, check these:

  • Is the licence OSI-approved (Apache 2.0, MIT, BSD)? If yes, standard open-source rules apply.
  • Does the licence include a commercial use grant? Be explicit — “research purposes only” means no commercial use.
  • Are there acceptable-use restrictions that could affect your use case?
  • Is redistribution of derived works (fine-tunes) explicitly allowed?
  • Is there a patent grant? Apache 2.0 includes one; custom licences may not.
  • Does the licence cap usage by organisation size, user count, or revenue?
  • Can the licence change for future versions, and would you need to re-check if you upgrade?

Methodology and sources

Check date: 2026-05-25

What was checked: Published model licences for Llama 2, Llama 3, Mistral 7B, Mixtral, Gemma, and several research-only models; OSI Open Source AI Definition (2024); legal commentary from software freedom organisations.

Assumptions and limits: This is not legal advice. Licence interpretation ultimately depends on your jurisdiction and use case. Model licences have been updated without public announcement — always check the current licence on the model’s official page before building.

Source list

Change Log

  • 2026-05-27: Added direct source URLs to all named providers and services; added Change Log section. Content unchanged.