Can I use this model commercially?

Usually yes, and the conditions are narrower than the paraphrased guides suggest. Models under Apache 2.0 or MIT ship commercially with notice conditions only. Llama 3.1 and later permit commercial use below 700 million monthly active users, plus naming and attribution rules. Gemma's terms state no user threshold at all. Qwen splits three ways, so the answer depends on which tier your exact model carries, and a model labeled non-commercial is not yours to sell.

This page settles the question in each license's own words, with the date each was read. Two corrections come with it, because most ranking pages still get them wrong: the training restriction that exists in Llama 2 and Llama 3 and is gone from Llama 3.1 onward, and the 200 million user limit that appears in no Gemma terms we could reach.

Not legal advice: read this first

The method, so you can check the page against itself: every quoted sentence is copied character for character from the linked text, each quote carries an as-of stamp, and where a license does not say something, this page says the license is silent rather than filling the gap.

The commercial-use answer, license by license

Apache 2.0 and MIT: yes, with paperwork. No threshold, no revenue line, no field of use. Apache 2.0 grants every recipient a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license" to reproduce, prepare derivative works of, and distribute the work (as of 2026-09-01). What the license asks in return is a copy of itself, notices on modified files, and any NOTICE file content carried along. MIT's entire condition is that the copyright and permission notice stays in copies. Most Mistral and Qwen releases live in this row.

Llama 3.1, 3.3 and 4: yes, below the threshold. The grant, quoted in full from the Llama 3.1 license (as of 2026-09-01):

"You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Meta’s intellectual property or other rights owned by Meta embodied in the Llama Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Llama Materials."

The conditions: stay under the 700 million monthly-active-user line (decoded below), comply with the Acceptable Use Policy the license incorporates, name models built on Llama with "Llama" at the beginning, display "Built with Llama", and carry the Notice text file when you distribute. The three version texts agree on all of this; only the version strings and dates differ. Llama 2 and Llama 3 granted the same commercial terms and added a training restriction that later versions dropped, which is the single most misquoted fact in this area.

Gemma: yes, under terms plus a policy. The Gemma Terms of Use permit commercial use and redistribution under four conditions, and the terms as of 2026-09-01 contain no monthly-active-user threshold in any version we could reach. They also state plainly that "Google claims no rights in Outputs you generate using Gemma." Gemma 4 models are pointed elsewhere: the terms' header reads "For Gemma 4 terms, see the Gemma 4 license", and that license is Apache 2.0.

Mistral: which model is it? Mistral 7B was released under Apache 2.0, and the Small, Devstral Small and Magistral Small releases we checked carried the apache-2.0 tag on Hugging Face as of 2026-09-01. The Large checkpoints are different: they sit under the Mistral Research License, which allows research and non-commercial use only, and commercial self-deployment needs their commercial license (both quoted in the table). A third tier, the Non-Production License, covers some tools and is non-commercial as well.

Qwen: three tiers, check the card. Most Qwen2.5 sizes and the Qwen3 releases carry Apache 2.0. Some releases carry the Qwen LICENSE AGREEMENT with a 100 million monthly-active-user threshold. Some carry the Qwen RESEARCH LICENSE AGREEMENT, which grants rights for non-commercial purposes only. The tier is printed on the model card, and the table quotes each one.

Anything labeled non-commercial: no. CC BY-NC and the research-only labels remove commercial use entirely. No threshold, no conditions to satisfy, no commercial product.

The five licenses that matter, in their own words

Quotes are elided with an explicit ellipsis where a sentence is cut; the full threshold clauses are quoted uncut in the next section. Meta's license texts are read from the LICENSE files in Meta's own llama-models repository (developer.meta.com serves the same text; Meta's ai.meta.com license URLs redirect onward to www.llama.com). The Apache 2.0 text is read from the copy mirrored in a GitHub repository.

The MAU thresholds, decoded

Two licenses in the table name a user threshold. Here is each clause in full, because the sentence does more work than the number usually quoted from it. The Llama 3.1 Community License, section 2, Additional Commercial Terms (as of 2026-09-01):

"If, on the Llama 3.1 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights."

The Qwen LICENSE AGREEMENT, section 4, Restrictions (as of 2026-09-01):

"If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, you shall request a license from us. You cannot exercise your rights under this Agreement without our express authorization."

What each sentence actually measures: Llama looks at the monthly active users of the products or services made available by or for the licensee and its affiliates, measured on the version release date, in the preceding calendar month. Qwen looks at your product or service. That is the whole of it. Neither license defines monthly active users, neither says how to count a business-to-business integration, and neither names an auditor. The license is silent beyond the sentence quoted, and this page will not invent counting rules the text does not contain.

Now the Gemma correction. The widely repeated figure is that Gemma carries a 200 million monthly-active-user limit. No version of the Gemma Terms of Use we could reach contains it: not the current text ("Last modified: April 1, 2026"), not the archived January 2025 text ("Last modified: April 1, 2024"), and not the launch text ("Last modified: February 21, 2024"). What the terms carry instead is the Prohibited Use Policy, "hereby incorporated by reference into this Agreement", and the clause that gives the terms their real teeth:

"To the maximum extent permitted by law, Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement."

The terms also state that "Google may update Gemma from time to time." A scale limit is a different mechanism from either of those, and only one of the two is a number.

What changed between Llama versions (and why most guides are wrong)

Llama 2 and Llama 3 both contain a sentence that stops you feeding Llama outputs into a competitor's model. From the Llama 3 license (as of 2026-09-01):

"You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Meta Llama 3 or derivative works thereof)."

Llama 2's text says the same with "Llama 2 or derivative works thereof" in the parentheses. From Llama 3.1 onward the sentence is gone. We fetched and searched all five version texts on 2026-09-01; the restriction is present in Llama 2 and Llama 3, and absent from Llama 3.1, 3.3 and 4.

This table is why paraphrased guides disagree with each other. A guide written from Llama 3 asserts a restriction that a guide written from Llama 3.1 will deny, and both call their model "Llama". One dated list we checked claims Llama 4 "adds further use restrictions"; we diffed the Llama 3.3 and Llama 4 license bodies and they are textually equivalent apart from version strings and the Effective versus Release date wording, so the claim is not supported by the license text. Use restrictions live in the Acceptable Use Policy, which each license incorporates by reference: your use must "adhere to the Acceptable Use Policy for the Llama Materials (available at https://llama.meta.com/llama3_1/use-policy), which is hereby incorporated by reference into this Agreement." Every quote on this page carries its date for exactly this reason.

Open weights are not open source

"Open weight" means you can download the trained model. It does not mean you got the recipe. A February 2025 study of openness in language models puts the frame in its own words: "The Open Source Initiative (OSI) has recently released its first formal definition of open-source software", and the authors "examine transparency and accessibility from two perspectives: open-source vs. open-weight models." Their finding applies directly to license pages: "while some models are labeled as open-source, this does not necessarily mean they are fully open-sourced." Even the best cases, they write, "often do not report model training data, and code as well as key metrics". The weights arrive; the recipe does not.

The Open Source Initiative publishes its own position on open weights at opensource.org; this page reaches it only through the study above. For a builder the definition debate does not change the practical rule: the license, not the open label, is the contract. A permissively licensed model with closed training data ships commercially; a restrictively licensed model with open tooling may not.

Redistribution: what ships with the weights

Sharing or re-serving a model makes you a distributor, and every license in the table names what a distributor carries. None of it is heavy. For Llama 3.1 and later, every copy keeps the Notice text file, which reads, verbatim: "Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved." Each version carries its own string: Llama 3's Notice names the Meta Llama 3 Community License, and Llama 2's reads "Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved." You also include "a copy of this Agreement", display "Built with Llama" (Llama 3 says "Built with Meta Llama 3"), and start derivative model names with the version mark. Fine-tunes are explicitly yours: "you are and will be the owner of such derivative works and modifications."

For Gemma, distributions carry a Notice text file reading "Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms", a copy of the agreement to each recipient, and the Section 3.2 use restrictions as an "enforceable provision in any agreement" you impose downstream. For Apache 2.0, recipients get "a copy of this License", modified files get prominent notices, and NOTICE file content travels along. For the Qwen tiers, recipients get a copy of the applicable agreement and models built with Qwen display “Built with Qwen” or “Improved using Qwen”.

One catch that trips people who fetch weights outside the official channel: some upstream repositories are gated. Meta's and Google's official model repositories require you to accept terms before downloading, while a fine-tune like NousResearch's Hermes can be ungated entirely. A torrent does not transfer terms you never accepted; the AI model torrents guide covers what that means for sharing by swarm.

A license tells you what you may do with a model. It says nothing about what a file will do when you load it; that question belongs to the model file safety guide.

Can I use it commercially? The decision block

Four questions, in order. Each one ends in an action, and the answer to the first decides which of the others matter.

1. Which license does this exact model carry? Read it from the model card, never from the marketing page or a blog summary. The identifier is in the card metadata, and the full text sits in a LICENSE file in the repository; the platform's own instruction is to "seek out and respect a project’s license if you’re considering using their code or data" (as of 2026-09-01). If the identifier reads other, open the LICENSE file before anything else.
2. Are you under the threshold it states, if it states one? Llama: 700 million monthly active users. Qwen's restricted tier: 100 million. Gemma, Apache 2.0 and MIT: no threshold as of this date. If you are above a stated threshold, the action is to request the license, and the Llama text is candid that Meta "may grant to you in its sole discretion".
3. Will you distribute, fine-tune, or re-share the weights? Then you carry the notice conditions in the redistribution section: notice file, license copy, attribution line, naming rule. The action is to put them in place before you ship, not after.
4. Is the model gated, and did you accept its terms yourself? If you never clicked accept on the upstream gate, the action is to read the terms before you build on the bytes. The obligation travels with the weights however you obtained them.

If your model is Apache 2.0 or MIT and you can meet the notice conditions, you are done: then only the notice conditions apply. No threshold to track, no permission to request, no policy incorporated by reference.

Licenses in this archive's catalog

The license label on every model page in this archive's catalog is the upstream record value, mirrored verbatim from the model's repository. It is recorded data, not our interpretation, and it is pinned to the exact upstream revision the payload was fetched from. Three worked examples:

• NousResearch_Hermes-3-Llama-3.1-8B carries the identifier llama3, which maps to the Llama 3 Community License Agreement: the version that still contains the training restriction, even though the model itself is a Llama 3.1 fine-tune. The upstream model card declares the same identifier.
• Qwen_Qwen2.5-3B-Instruct carries the identifier other. Its upstream card sets license_name to qwen-research, the non-commercial tier quoted in the table. The identifier table's instruction for this case is "In case of license: other please add the license’s text to a LICENSE file inside your repo", which is where we read it. The site's non-commercial badge only matches identifiers containing "-nc-", so no warning badge renders on that model page today; read the license line, not the badge.
• Qwen_Qwen2.5-7B-Instruct carries apache-2.0, read straight from the card.

A label is a lookup key, not the license. The bytes being correct and the terms being acceptable are two different checks, and only the first one is ours. Verification in this archive means every payload was verified against upstream Hugging Face at fetch time, with the exact upstream revision pinned in the signed manifest, using sha256 for LFS files and sha1+size for git blobs, each labeled with its method. That is the whole claim. Seedbank does not audit model behavior. How to run the check is on the verify page and in the verification guide, and whether the platform hosting a model is trustworthy at all is a separate question covered in are Hugging Face models safe.