Skip to content

Big-Tech Model Licenses in Practice: A Side-by-Side Comparison of Seven Companies ​

We · October 2026

This is the sixth article in our "Open Source and Commercial Use" series. The previous five covered why companies open-source their models (Why Companies Open Source), how software licenses get swapped (Why Open-Source Projects Change Licenses Mid-Stream), what happens to license obligations after you modify the weights (Do Quantized, LoRA, or Fine-Tuned Models Still Fall Under the Original License?), how schools and research labs should handle licenses (Do Schools and Research Labs Need to Worry About Licenses?), and why people who aren't trying to make money still open-source (Why People Who Don't Run a Business Also Open-Source). This article covers what big tech actually wrote — every clause below comes from the LICENSE file, not from rumor.

For the rules governing software licenses themselves, see Open-Source Software Licenses Explained. For whether each downloadable model on our platform can be used commercially, see audio.cpp Model Licenses and Commercial Use.

All clause descriptions in this article are sourced from each repository's LICENSE file and public reporting, compiled as of October 2026. We do not review, rank, link to external sites, or provide legal advice — always refer to each repository's LICENSE file at the time of use.

I. What This Article Covers — and What It Doesn't ​

The previous five articles in this series covered motivations, clause design, usage scenarios, and publisher structures. This one covers what big tech actually wrote.

Seven companies, seven licenses, each with a different clause design. Some set scale thresholds, others set revenue caps, some require attribution, some reserve revocation rights, and one splits code and weights into two separate licenses. None of these clauses are hidden — they are all in the LICENSE file. The reason this article is worth writing is simply that most people never open that file.

This article does not cover software licenses (that is Open-Source Software Licenses Explained), does not go model-by-model on commercial use (that is audio.cpp Model Licenses and Commercial Use), and does not cover why companies open-source (that is the first article). It does only one thing: lays out seven companies' clauses side by side.

II. Seven Companies' Clause Design: A Side-by-Side Comparison Table ​

The table below is organized by clause type, not ranked, and not a recommendation order. Each cell represents a design difference in clause terms — always verify against the original LICENSE file for specific projects.

Company / ProjectScale ThresholdRevenue CapAttribution ObligationNo Training Competing ModelsRevocabilityGeographic RestrictionSeparate Code and Weight Licenses
Meta Llama700M MAU (introduced with Llama 2, retained in 3.1/3.3/4)NoneDerivative models from Llama 3.1 onward must prominently attribute "Llama"; Llama 4 requires displaying "Built with Llama"Llama 2/3 prohibit using model outputs to train other models; from 3.1 onward this is allowed but requires attributionLlama 4 explicitly states Meta may terminate the licenseLlama 4 multimodal models have distribution restrictions for EU-based enterprisesNo (unified community license)
Alibaba Qwen100M MAU (custom Qwen License)None"Built with Qwen" attribution clause; disputes governed by Chinese law under Hangzhou court jurisdictionNo explicit clauseYes (custom license reserves termination rights)No explicit clauseNo (unified custom license)
DeepSeekNoneNoneNoneNoneNone (MIT license is irrevocable)NoneNo (unified MIT)
Stability AINoneFree commercial use for annual revenue under $1M; above that requires enterprise license and registrationNo explicit clauseMay not use outputs to train competing foundation modelsYes (community license reserves termination rights)No explicit clauseNo (unified community license)
Black Forest Labs FLUXNoneNoneNoneNoneschnell: No (Apache 2.0 is irrevocable); dev: Yes (research license)NoneYes (schnell Apache 2.0, dev non-commercial research license — weights in the same repo under different files with different licenses)
Google GemmaNoneNoneNoneNoneYes (custom license reserves termination rights)NoneNo (unified custom license); includes a Prohibited Use list (weapons, biohazards, and other misuse categories)
MistralNoneNoneNoneNone7B: No (Apache 2.0 is irrevocable); flagship: Yes (research license)NoneNo (7B Apache 2.0, flagship research license — dual-track)

Three things stand out from this table:

First, scale thresholds are a big-tech design. Llama's 700M MAU and Qwen's 100M MAU are gates built for large-scale service providers. Small users are completely free; large users must negotiate a commercial license. Stability's revenue line follows the same logic, just drawn on money instead of headcount.

Second, revocability is the dividing line between permissive and custom licenses. Rights granted under MIT and Apache 2.0 are irrevocable — once given, the author cannot take them back. Custom community licenses from various publishers universally reserve termination rights. Permissive licenses give you more freedom; custom licenses give the publisher more power to take it back. This is not a coincidence — it is design.

Third, separate licensing for code and weights is the norm, not the exception. FLUX is the most典型 example: schnell is Apache 2.0, dev is a non-commercial research license, and both sets of weights sit in the same repository under different files with different licenses. Different generations within the same product line can also carry completely different licenses — Qwen went from research license to Apache 2.0 to a custom Qwen License; Qwen-Image's previous generation was Apache 2.0, and version 2.1 switched back to a research license. Always read the file, not the label.

III. Three Cases Worth Examining Separately ​

Case 1: Llama 4's Trademark Tie-In and Revocability ​

The Llama 4 community license (effective April 5, 2025) pushed clause design to a new level:

  • Mandatory brand display: Redistribution or derivative products must display "Built with Llama."
  • Name prefix obligation: Derivative models trained on Llama outputs must have names starting with "Llama."
  • Geographic restriction: Distribution of multimodal models carries restrictions for EU-based enterprises.
  • Revocability: Meta explicitly reserves the right to terminate the license.
  • Clause relaxation: Reportedly removed the Llama 2/3 clause prohibiting development of competing products.

The message of this combination is clear: you can use it for free, but you have to let me know you're using it, and I have to be able to take it back at any time. This is a completely different philosophy from MIT's "once given, never taken back."

Case 2: Qwen-Image 2.1's Re-License Controversy ​

Qwen-Image's previous generation was Apache 2.0. Version 2.1, released in September 2026, switched back to a research license — SaaS services and secondary distribution of weights now require a commercial license application. This re-license sparked a public debate in the technical community about "where is the open-source promise."

This case is noteworthy not because of the clause itself (re-licensing is generally legally permissible), but because the cost of re-licensing is not in the announcement — it is in trust. The second article in this series, Why Open-Source Projects Change Licenses Mid-Stream, covered the re-license histories of Elasticsearch and Redis: changing back cannot undo a fork. Qwen-Image has not reached that point yet, but the debate itself demonstrates: users' expectations of "open weights" are stable, and clause oscillation erodes that expectation.

Case 3: Stability's Registration Requirement ​

Stability AI's community license states: free commercial use for annual revenue under $1M; above that requires purchasing an enterprise license and registering on the official website for it to take effect.

That last phrase is the one most easily overlooked. Without registration, the license itself may not be valid. The revenue line is pricing; registration is the condition for effectiveness — the two stacked together form the complete contract. This is also why the first article in this series listed "registration required for effectiveness" as one of nine hidden details: no clause is hidden in the dark, it's just that most people never read the file.

IV. Four Practical Recommendations ​

First, when you see a big-tech model, look for the scale threshold and revenue line first. Llama's 700M MAU, Qwen's 100M MAU, Stability's $1M revenue line — these numbers determine whether you are a free user or one who needs to negotiate a contract. Most individual developers and small teams fall below the threshold and are completely free; but if you are running a service, raising funds, or preparing to scale, that one line is worth checking early.

Second, when you see "Built with X" or a name prefix obligation, recognize it as a brand clause, not a technical restriction. Both Llama 4 and the Qwen License require derivative products to display the brand or name prefix. The purpose of such clauses is to ensure the publisher's name travels with your product. You can accept it or not, but do not mistake it for a technical constraint.

Third, when code and weights are split across different files, verify each one's license separately. FLUX is the most typical example: schnell is Apache 2.0, dev is a research license, and both sit in the same repository. GitHub's label may show only the primary license, but in practice you need to look at the LICENSE file next to the specific weight file you downloaded. Labels can be wrong; files are not.

Fourth, when you see a custom community license, assume it has revocation rights. MIT and Apache 2.0 are irrevocable — this is the most substantive value of permissive licenses. Custom licenses (Llama Community License, Qwen License, Stability Community License, Gemma License) universally reserve termination rights. This is not "might have" revocation — it is "definitely has" it. If your business depends heavily on a particular model, a custom license means you are standing on an authorization that can be withdrawn at any time.

V. Our Take ​

First, clause design reflects the publisher's understanding of the business. Llama wants ecosystem positioning, so it sets an MAU threshold — small users are free, large users negotiate contracts. Qwen wants developer adoption, so it sets a 100M MAU threshold and attribution obligation — if you use my model, I want to know about it. Stability wants revenue, so it sets a $1M line — below it is free for acquisition, above it is paid. No license is "standard" — each one is custom-built for a specific business model.

Second, the difference between permissive and custom licenses is not "can you use it" but "can it be taken back after you use it." MIT and Apache 2.0 give you an irrevocable authorization on which you can build long-term plans. Custom licenses give you "temporary permission" — the publisher can change terms or revoke access at any time. If your business depends on a particular model, this difference matters more than the model's capabilities.

Third, re-licensing is the norm, not the exception. Qwen-Image 2.1 switched from Apache back to a research license; Mistral's flagship went from a research license back to Apache 2.0 — the direction is not consistently tightening or consistently loosening, but oscillating with the company's business stage. This is why "pinning versions" is genuinely effective in a legal sense, not just an operational habit. Record which version you are using, and record the license name alongside it.

With this article, the big-tech side of clause design is complete.

Disclaimer: This article is our compilation of each repository's LICENSE file and public reporting. It does not constitute legal advice. Clause descriptions are current as of October 2026 and may change with upstream updates. For any decision involving real commercial interests, consult professional legal counsel and always refer to the project's current LICENSE file.