Model Licenses Explained: Research Licenses, Community Licenses, Custom Licenses ​

Written: October 2026. License terms follow each project's official version. This article is for informational purposes only and does not constitute legal advice.

When you download model weights from Hugging Face, you'll likely see license names like:

• Llama 3.2 Community License
• Qwen Research License
• Stability AI Community License
• FLUX.1 [dev] Non-Commercial License
• Gemma Terms of Use

These are not OSI (Open Source Initiative) approved open source licenses. They're not MIT, Apache-2.0, GPL, nor part of the Creative Commons family. They're custom licenses — drafted individually by model publishers, designed specifically for those models.

Series 13 #1 mentions "open weights ≠ open source," and #5 compares seven companies' clause designs side by side. This article systematically covers the common features, typical clauses, and fundamental differences from OSI open source licenses in these "non-standard" licenses.

1. Why Don't Model Weights Use Standard Open Source Licenses? ​

1.1 Open Source Licenses Were Designed for Code ​

MIT, Apache, GPL — these licenses were drafted with source code in mind: text files, compilable, diffable, mergeable. Their core concepts — "source code," "compilation," "linking," "distribution" — all revolve around software code.

Model weights aren't code. They are:

• Binary files: Tens of gigabytes of floating-point matrices, not human-readable, not diffable
• Training artifacts: Not written by humans, but learned from data
• Dependent on training data: The nature of weights is influenced by training data (if training data contains copyrighted content, the weights themselves may have copyright issues)
• Quantizable, fine-tunable, mergeable: A weight file can be converted, derived into multiple versions

Using software licenses for model weights creates legal interpretation ambiguity.

1.2 Model Publishers Need Finer Control ​

When big tech releases model weights, they usually have commercial considerations:

• Meta: Wants the Llama ecosystem to grow, but doesn't want others using Llama to train competing models
• Alibaba: Wants Qwen widely adopted, but over 100 million MAU requires separate licensing
• Stability AI: Wants Stable Diffusion used by creators, but annual revenue over $1 million requires payment
• Black Forest Labs: FLUX.1 [schnell] is fully open source (Apache 2.0), but FLUX.1 [dev] only allows non-commercial research

These commercial strategies can't be expressed in standard open source licenses. MIT doesn't let you add "MAU caps," Apache doesn't let you add "revenue thresholds," GPL doesn't let you add "no training competing models."

So big tech chose to write their own licenses.

1.3 "Open Weights" ≠ "Open Source" ​

The Open Source Initiative (OSI) has a clear definition of "open source": must allow free use, modification, distribution, without discriminatory restrictions. Many model weight licenses don't meet this definition:

• Llama Community License: MAU over 700 million requires separate licensing → has thresholds
• Qwen License: MAU over 100 million requires separate licensing → has thresholds
• Stability AI Community License: Annual revenue over $1 million requires payment → has thresholds
• FLUX.1 [dev] License: Prohibits commercial use → has discriminatory restrictions
• Gemma Terms of Use: Prohibits certain specific uses (Prohibited Use) → has discriminatory restrictions

These licenses are more accurately called "open-weight" or "source-available," not "open source."

Series 13 #1 explains the conceptual distinction between "open weights" and "open source" in detail; this article focuses on the specific clauses in these non-standard licenses.

2. Common Features of Non-Standard Model Licenses ​

Although each company's license wording differs, they share some common features:

2.1 Usage Thresholds ​

Most non-standard licenses set usage thresholds, above which separate licensing or payment is required:

How thresholds are calculated:

• MAU (Monthly Active Users): Usually refers to "independent users who used the model or its derivatives in the past 30 days"
• Revenue: Usually refers to "total revenue earned in the past 12 months from using the model or its derivatives"

Note: Thresholds are usually global, not calculated by region or product line. If your company has 800 million MAU globally, and you use a Llama model to build an internal tool used by only 1,000 people, you still exceed the 700 million MAU threshold.

2.2 Non-Compete Clauses ​

Many non-standard licenses prohibit you from using the model or its outputs to train competing models:

• Llama Community License: Prohibits using Llama model outputs (including model-generated text, images) to train other large language models
• Qwen License: Prohibits using Qwen models to train models competing with Qwen
• Stability AI Community License: Prohibits using Stability models to train models competing with Stability

What counts as "competing"? Licenses usually don't define it. But generally understood as:

• Same type of model: Using Llama (language model) outputs to train another language model counts as competing
• Same function model: Using Stable Diffusion (image generation) outputs to train another image generation model counts as competing
• Cross-type model: Using Llama outputs to train an image generation model might not count as competing (disputed)

What counts as "training"? Usually includes:

• Full parameter fine-tuning: Yes
• LoRA fine-tuning: Yes
• Distillation: Yes
• Using model outputs for data labeling, then training a new model: Yes (this is the most common workaround, but most licenses explicitly prohibit it)

2.3 Trademark Restrictions ​

Many non-standard licenses prohibit you from using the publisher's trademarks:

• Llama 4 License: Prohibits using "Meta" "Llama" trademarks for promotion
• Qwen License: Prohibits using "Alibaba" "Qwen" "Tongyi" trademarks for promotion
• Gemma Terms of Use: Prohibits using "Google" "Gemma" trademarks for promotion

Exception: Some licenses allow you to use trademarks when stating product compatibility. For example, you can write "This product is compatible with Llama 3.2," but can't write "This product is supported by Meta."

2.4 Geographic Restrictions ​

A few non-standard licenses have geographic restrictions:

• Some Chinese companies' licenses: May restrict use in certain countries or regions
• Some US companies' licenses: May restrict use in sanctioned countries

Note: Geographic restrictions are usually tied to export control laws, not specified by the license alone.

2.5 Revocability ​

Most non-standard licenses are revocable:

• Llama Community License: Meta reserves the right to revoke the license at any time
• Qwen License: Alibaba reserves the right to revoke the license at any time
• Stability AI Community License: Stability reserves the right to revoke the license at any time

Consequences of revocation: If the license is revoked, you must stop using the model and destroy all copies. This is much riskier than irrevocable licenses like MIT, Apache.

Why design licenses as revocable? Because model publishers want to retain control. If you violate the license terms (like exceeding the MAU threshold without applying for authorization), the publisher can revoke your license.

2.6 Disclaimer ​

All non-standard licenses have strong disclaimers:

• "The model is provided as-is, without any express or implied warranties"
• "The publisher is not responsible for the accuracy, reliability, or legality of model outputs"
• "All risks from using the model are borne by the user"

These clauses exist in standard licenses like MIT, Apache too, but non-standard licenses usually write them in more detail, with more emphasis.

3. Common Non-Standard Licenses in Detail ​

3.1 Llama Community License ​

Applicable models: Llama 2, Llama 3, Llama 3.1, Llama 3.2, Llama 4

Core clauses:

• MAU threshold: Monthly active users over 700 million requires separate licensing
• Non-compete: Prohibits using Llama model outputs to train other large language models
• Trademark restriction: Prohibits using "Meta" "Llama" trademarks for promotion
• Revocability: Meta reserves the right to revoke the license at any time
• Disclaimer: Model provided as-is, no warranties

Llama 4 special clauses:

• Trademark tie-in: Llama 4 license prohibits using "Llama" in product names, domains, trademarks
• Enhanced revocability: If you violate the license terms, Meta can immediately revoke your license

Practical advice:

• If your product MAU doesn't exceed 700 million, you can use Llama freely
• Don't use Llama outputs to train competing models
• Don't use "Llama" in product names (like "My Llama Assistant")

3.2 Qwen License ​

Applicable models: Qwen, Qwen 1.5, Qwen 2, Qwen 2.5, Qwen 3

Core clauses:

• MAU threshold: Monthly active users over 100 million requires separate licensing
• Non-compete: Prohibits using Qwen models to train models competing with Qwen
• Trademark restriction: Prohibits using "Alibaba" "Qwen" "Tongyi" trademarks for promotion
• Revocability: Alibaba reserves the right to revoke the license at any time
• Disclaimer: Model provided as-is, no warranties

Qwen-Image 2.1 re-licensing controversy:

• Qwen-Image 2.1 originally used Apache 2.0, later changed to Qwen License
• This caused projects already developed based on the Apache 2.0 version to face stricter restrictions in the new version
• Series 13 #5 covers this case in detail

Practical advice:

• If your product MAU doesn't exceed 100 million, you can use Qwen freely
• Don't use Qwen to train competing models
• Note that different Qwen versions may have different licenses

3.3 Stability AI Community License ​

Applicable models: Stable Diffusion 3, Stable Audio, Stable Video

Core clauses:

• Revenue threshold: Annual revenue over $1 million requires payment
• Non-compete: Prohibits using Stability models to train models competing with Stability
• Trademark restriction: Prohibits using "Stability AI" "Stable Diffusion" trademarks for promotion
• Revocability: Stability reserves the right to revoke the license at any time
• Disclaimer: Model provided as-is, no warranties

Practical advice:

• If your company's annual revenue doesn't exceed $1 million, you can use Stability models freely
• Don't use Stability models to train competing models
• Note how "revenue" is calculated: all revenue, or only revenue generated from using the model? Licenses are usually unclear

3.4 FLUX.1 License (Dual License Model) ​

Applicable models: FLUX.1 [schnell], FLUX.1 [dev], FLUX.1 [pro]

Core clauses:

• FLUX.1 [schnell]: Apache 2.0, fully open source, commercial use allowed
• FLUX.1 [dev]: Non-commercial research license, commercial use prohibited
• FLUX.1 [pro]: Commercial license, requires payment

Dual license strategy:

• Black Forest Labs uses FLUX.1 [schnell] (Apache 2.0) to attract the community
• Uses FLUX.1 [dev] (non-commercial) to restrict competitors
• Uses FLUX.1 [pro] (paid) to generate revenue

Practical advice:

• If you want to commercially use FLUX models, you can only use FLUX.1 [schnell]
• FLUX.1 [dev] can only be used for personal learning, research
• FLUX.1 [pro] requires contacting Black Forest Labs for a commercial license

3.5 Gemma Terms of Use ​

Applicable models: Gemma, Gemma 2, Gemma 3

Core clauses:

• Prohibited Use: Prohibits using Gemma for:
  • Generating misinformation
  • Generating hate speech
  • Military use
  • Surveillance or privacy violations
• Trademark restriction: Prohibits using "Google" "Gemma" trademarks for promotion
• Revocability: Google reserves the right to revoke the license at any time
• Disclaimer: Model provided as-is, no warranties

Prohibited Use boundaries:

• What counts as "misinformation"? The license doesn't define it. Generally understood as "intentionally spread false content"
• What counts as "hate speech"? The license doesn't define it. Generally understood as "discrimination, violence, hate speech against specific groups"
• What counts as "military use"? The license doesn't define it. Generally understood as "direct use in military operations or weapons systems"

Practical advice:

• Gemma has stricter usage restrictions than other models
• If your application involves content generation, you need to carefully review the Prohibited Use clauses
• Don't use "Gemma" in product names

4. Fundamental Differences Between Non-Standard and OSI Open Source Licenses ​

Fundamental difference:

• OSI open source licenses center on freedom: you can freely use, modify, distribute, as long as you follow a few obligations (attribution, viral sharing, etc.)
• Non-standard model licenses center on control: publishers want to retain control over model usage, including who can use it, how, and for what

5. How to Read a Non-Standard License? ​

When you get a non-standard license, read in this order:

5.1 First Look at "Grant" ​

• What rights does the license grant you? Use? Modify? Distribute? Commercial use?
• Are these rights conditional?

5.2 Then Look at "Restrictions" ​

• What behaviors are explicitly prohibited?
• Prohibited from training competitors? Prohibited from commercial use? Prohibited from certain specific uses?

5.3 Look at "Threshold" ​

• Are there MAU, revenue, or other thresholds?
• What's required above the threshold? Separate licensing? Payment?

5.4 Look at "Trademark" ​

• Are you allowed to use the publisher's trademarks?
• Under what circumstances can you use them?

5.5 Look at "Revocability" ​

• Is the license revocable?
• What are the consequences of revocation?

5.6 Look at "Disclaimer" ​

• Does the publisher provide any warranties?
• Who bears the risk of using the model?

5.7 Look at "Termination" ​

• Under what circumstances does the license terminate?
• What must you do after termination?

6. Common Non-Standard Licenses in Local AI Apps ​

From the 94 local AI apps we've cataloged, these use non-standard licenses:

Note: The same app may load multiple models, each with different licenses. For example, Ollama can load Llama (Llama Community License), Qwen (Qwen License), Mistral (Apache 2.0), and other models.

7. Common Misconceptions ​

Misconception 1: Non-standard licenses are "open source" ​

Wrong. Non-standard licenses usually have thresholds and restrictions, not meeting OSI's definition of "open source." More accurate terms are "open-weight" or "source-available."

Misconception 2: You only need to follow the license above the MAU threshold ​

Wrong. The MAU threshold is just the point where separate licensing is needed, not where license obligations begin. Even if your product has only 1,000 MAU, you still need to follow other license terms (like non-compete, trademark restrictions, etc.).

Misconception 3: Non-standard licenses are irrevocable ​

Wrong. Most non-standard licenses are revocable. If the publisher revokes your license, you must stop using the model. This is much riskier than irrevocable licenses like MIT, Apache.

Misconception 4: Prohibited Use in non-standard licenses is just a "suggestion" ​

Wrong. Prohibited Use is a legal constraint. If you violate it, the publisher can revoke your license, or even sue you.

Misconception 5: Non-standard licenses are compatible with OSI open source licenses ​

Not necessarily compatible. For example, GPL requires "no additional restrictions," but non-standard licenses usually add MAU thresholds, non-compete restrictions. If you merge GPL code with non-standard licensed model weights into one project, there may be legal conflicts.

8. Relationship to Other Articles ​

• Series 13 #1 (Why Commercial Companies Open Source): Covers the concept of "open weights ≠ open source." This article details the specific clauses in non-standard licenses.
• Series 13 #5 (Big-Tech Model Licenses in Practice): Side-by-side comparison of seven companies' licenses. This article systematically covers the common features of these licenses.
• Series 13 #8 (Derivative Model License Chains): Covers whether quantized, LoRA, fine-tuned models are subject to original licenses. This article's "non-compete" clauses are relevant.
• skill/licensing/open-source-software-licenses: Covers standard open source licenses like MIT, Apache, GPL. This article covers non-standard model licenses. The two complement each other.

9. References ​

• Llama Community License Full Text
• Qwen License Full Text
• Stability AI Community License Full Text
• FLUX.1 License Full Text
• Gemma Terms of Use Full Text
• OSI Definition of "Open Source"
• Hugging Face Model License Guide

Disclaimer: This article is for informational purposes only and does not constitute legal advice. For specific compliance questions, consult a professional lawyer. License terms follow each project's official LICENSE file; translations are not authoritative.