Best Open Source AI

Open source vs open weights: what each licence lets a company do

Almost every model sold as open-source AI publishes its weights but not its training data. The difference rarely matters for internal use. It matters a great deal for companies that resell a model or build a product on it.

By The editors · Updated · 6 min read

A legal counsel and a product engineer read a printed licence agreement together at a meeting table, one marking a clause with a pen.
The licence, not the model card, decides what a company may do with a model.
Contents
  1. Why the word matters
  2. What the definition requires
  3. The licences, family by family
  4. What this means in practice
  5. The EU's own definition
  6. How to choose
  7. Questions buyers ask
  8. Sources

Why the word matters

Companies searching for open-source AI are usually looking for three things: a model they can run themselves, freedom from per-token pricing, and the right to adapt it. Most models labelled open source give them all three. What they rarely give is what the term means in software: the full means of reproducing the work. The distinction is easy to ignore until a company wants to build a product on a model, offer it to customers, or answer a regulator's question about what the model was trained on.

What the definition requires

The Open Source Initiative's Open Source AI Definition 1.0 extends the four software freedoms (to use, study, modify and share) to AI systems. To qualify, three components must be available under approved terms⁠1:

  • Data information: sufficiently detailed information about the data used to train the system.
  • Code: the complete source code used to train and run the system.
  • Parameters: the model weights.

By that standard, most of the market is open-weight. Of the families reviewed on this site, only Ai2's OLMo publishes its pretraining and post-training data, its training code and its intermediate checkpoints, under Apache 2.0⁠23. Google describes Gemma 4's training data but does not release it⁠4, and the other families publish weights and usage information only.

The licences, family by family

The more practical question is what each licence allows. The current licences fall into two groups: standard permissive licences, and custom licences with conditions that mostly apply at scale or when the model is resold.

ModelLicenceConditions that matter to a company
DeepSeek V4-Pro, V4.1-FlashMITNone beyond the licence notice⁠5
Gemma 4Apache 2.0None beyond the licence notice⁠6
gpt-oss-120b and 20bApache 2.0 plus a usage policyUse must comply with applicable law⁠7
OLMo 3.1, Olmo-HybridApache 2.0None beyond the licence notice⁠2
Mistral Small 4, Ministral 3Apache 2.0None beyond the licence notice⁠8
Qwen3.8-27BApache 2.0None beyond the licence notice⁠9
GLM-5.3-Flash, GLM-5.2MITNone beyond the licence notice⁠1011
GLM-5.3GLM-5.3 License (MIT-style)A model-as-a-service operator with more than $10 billion in group revenue in any 12 months must pass a security review by Z.ai⁠12
Qwen3.8-2.4T-A95BQwen3.8-Max LicenseThe model name must be displayed above 100 million monthly users or $20 million monthly revenue; model-as-a-service and AI-assistant businesses above $50 million revenue need a separate licence; internal use is exempt⁠13
Mistral Large 4Mistral Ecosystem Open-Weight License (modified MIT)Offering it as a model-as-a-service needs a separate commercial licence; internal use is exempt⁠14
Muse Spark OpenMuse Model LicenseA "Built with Muse Spark Open" notice; service operators above $20 million revenue need a separate agreement; a usage policy applies⁠15
Llama 4 Maverick, ScoutLlama 4 Community LicenseA licence request above 700 million monthly users, a "Built with Llama" notice, "Llama" at the start of derivative model names and an acceptable use policy⁠16

Two patterns stand out. First, the restrictive terms sit on the largest flagships, while the same developers release smaller models under MIT or Apache 2.0: Qwen3.8-27B, GLM-5.3-Flash, Mistral Small 4 and Meta's Muse Glimmer 30B, which is Apache 2.0⁠9101718. Second, most conditions target companies that sell access to the model, not companies that use it inside their own operations.

What this means in practice

Internal use. A company running a model for its own staff and processes faces few constraints under any licence above. The Qwen and Mistral custom licences explicitly exempt internal use from their service conditions⁠1314.

Building a product on it. A company that puts a model inside a product sold to customers should read the notice and naming duties. Llama 4 requires a "Built with Llama" notice, and Muse Spark Open requires a "Built with Muse Spark Open" notice⁠1615.

Offering it as a service. Reselling access to a model is where the custom licences bite: the separate licence for Mistral Large 4, the revenue thresholds for Qwen3.8-Max and Muse Spark Open, and the security review for the largest GLM-5.3 operators⁠14131512.

Fine-tuned models. Fine-tuning creates a derivative, and custom licences can carry their conditions over to it. Llama 4's rule that derivative model names begin with "Llama" is the clearest example⁠16. A company that plans to train a model on its own data and keep it should start from a permissive licence where possible; our guide to training a model on company data sets out the steps.

The EU's own definition

The AI Act uses its own test. Models released under a free and open-source licence with public weights, architecture and usage information are exempt from part of the documentation duties for general-purpose models, but not from the copyright policy or the public summary of training content, and not at all if the model has systemic risk⁠1920. The Commission's conditions also exclude monetised releases⁠21, so a custom licence that charges service operators may not qualify. Our guide to the AI Act and open-source models explains what this means for a company that fine-tunes one.

How to choose

  1. Decide how the model will be used: internally, inside a product, or resold as a service.
  2. Read the licence of the exact model, not the family. Sizes within one family often carry different licences.
  3. Prefer MIT or Apache 2.0 for anything the company plans to fine-tune and keep, so the result carries no naming or revenue conditions.
  4. Choose a fully open model such as OLMo when the company must be able to show what the model was trained on.
  5. Record the licence and version of every model in use, and re-check when a new version is adopted.

A company that follows these steps knows, for every model behind its router, what it may do with it, what it must display and what would change if it began selling access.

Questions buyers ask

What is the difference between open source and open weights?

An open-weight model publishes its trained parameters so anyone can download and run it. An open-source AI system, as the Open Source Initiative defines it, must also make available sufficiently detailed information about its training data and the complete code used to train and run it. Most models called open source publish weights only.

Which AI models are truly open source?

Of the families reviewed on this site, only Ai2's OLMo publishes its training data, training code and intermediate checkpoints alongside its weights under Apache 2.0, which is the closest match to the OSI definition. The others, including DeepSeek, Gemma, gpt-oss, Qwen, Mistral, GLM and Meta's models, are open-weight.

Can I use an open-weight model commercially?

For most current models, yes. MIT and Apache 2.0 models such as DeepSeek V4, Gemma 4, gpt-oss, Mistral Small 4 and OLMo allow commercial use, modification and redistribution. Custom licences add conditions such as naming rules, revenue thresholds or a separate agreement to offer the model as a service.

Is Llama open source?

Not under the OSI definition. Llama 4 uses Meta's community licence, which requires a separate licence above 700 million monthly active users, a "Built with Llama" notice and "Llama" at the start of the names of derivative models, and an acceptable use policy. Meta's newer Muse Spark Open model has its own licence with similar branding duties.

Do I need a special licence to fine-tune an open model?

Not for the MIT and Apache 2.0 models, which allow modification. Custom licences also generally permit fine-tuning but may carry their conditions over to the result, such as naming rules for derivative models. The conditions that most often bite apply when the model or a derivative is offered to other companies as a service.

Does the EU AI Act treat open-source models differently?

Yes. Models released under a free and open-source licence, with public weights and architecture, are exempt from some documentation duties, but not from the copyright policy or the public training-data summary, and the exemption does not apply to models with systemic risk.

Sources

  1. The Open Source AI Definition 1.0. Open Source Initiative. Accessed 9 Nov 2026.
  2. allenai/Olmo-3.1-32B-Think. Ai2 (Hugging Face). Accessed 9 Nov 2026.
  3. allenai/Olmo-Hybrid-7B. Ai2 (Hugging Face). Accessed 9 Nov 2026.
  4. google/gemma-4-31B-it. Google (Hugging Face). Accessed 9 Nov 2026.
  5. deepseek-ai/DeepSeek-V4.1-Flash. DeepSeek (Hugging Face). Accessed 9 Nov 2026.
  6. Gemma 4 licence (Apache License 2.0). Google AI for Developers. Accessed 9 Nov 2026.
  7. gpt-oss usage policy. OpenAI (Hugging Face). Accessed 9 Nov 2026.
  8. Models overview. Mistral AI Docs. Accessed 9 Nov 2026.
  9. Qwen/Qwen3.8-27B. Qwen (Hugging Face). Accessed 9 Nov 2026.
  10. zai-org/GLM-5.3-Flash. Z.ai (Hugging Face). Accessed 9 Nov 2026.
  11. zai-org/GLM-5.2. Z.ai (Hugging Face). Accessed 9 Nov 2026.
  12. GLM-5.3 License. Z.ai (Hugging Face). Accessed 9 Nov 2026.
  13. Qwen3.8-Max License. Qwen (Hugging Face). Accessed 9 Nov 2026.
  14. Mistral Ecosystem Open-Weight License. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
  15. meta-models/Muse-Spark-Open-816B-A42B. Meta (Hugging Face), 15 Oct 2026. Accessed 9 Nov 2026.
  16. Llama 4 Community License Agreement. Meta (Hugging Face), 5 Apr 2025. Accessed 9 Nov 2026.
  17. mistralai/Mistral-Small-4-119B-2603. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
  18. meta-models/Muse-Glimmer-30B. Meta (Hugging Face). Accessed 9 Nov 2026.
  19. Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex, 12 Jul 2024. Accessed 9 Nov 2026.
  20. Guidelines for providers of general-purpose AI models. European Commission. Accessed 9 Nov 2026.
  21. Questions and answers on the guidelines for general-purpose AI providers. European Commission. Accessed 9 Nov 2026.

Updated 9 Nov 2026: First published. Licences read on the model cards and licence files on 9 November 2026.

Further reading

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski