Open-Weight Is Not Open-Source
What "Open" AI Models Actually Let You Do

When an AI company says it has released a model “openly”, it often means you can download the trained model and run it yourself. That is useful. It is also not the same thing as open-source — and the difference matters if you want to build on the model, study how it was made, or use it in a business.

This page explains the difference and shows what the licences on several downloadable models say, as read in September 2026. It is an explainer, not legal advice; if you plan to use a model commercially, read its licence in full.

What “weights” are

A trained AI model is, at its core, a very large set of numbers learned during training. The Open Source Initiative (OSI), which published the definition of open-source AI discussed below, explains: “Open Weights refer to the final weights and biases of a trained neural network.”

Having those numbers lets you run the model. It does not tell you how the model was built. As the OSI puts it: “However, Open Weights differ significantly from Open Source AI because they do not include:” the training code, the training dataset and information about the data.

That still gives you real options. In a September 2026 post, the OSI wrote that with open weights “you can choose to run a model locally” and “You can also fine-tune the model with your own data”. It also set out the limit: “Open-weights allow you to exercise some freedoms, but not all, and only to a certain degree.”

What “open-source AI” requires under the OSI definition

The OSI published version 1.0 of its Open Source AI Definition in October 2024; its announcement is datelined “RALEIGH, N.C., Oct. 28, 2024”. It starts from four freedoms. An open-source AI system must let you:

1. “Use the system for any purpose and without having to ask for permission.”
2. “Study how the system works and inspect its components.”
3. “Modify the system for any purpose, including to change its output.”
4. “Share the system for others to use with or without modifications, for any purpose.”

To make those freedoms real, the definition lists three things that must be available:

• Data information — “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.” It does not require every piece of training data to be published. It requires a complete description of the data, “including (if used) of unshareable data”, plus “a listing of all publicly available training data and where to obtain it” and “a listing of all training data obtainable from third parties and where to obtain it, including for fee”.
• Code — “The complete source code used to train and run the system.”
• Parameters — “The model parameters, such as weights or other configuration settings.” And: “Parameters shall be made available under OSI-approved terms.”

So weights are one of three required parts. Under this definition, a model whose weights are released under a standard licence is still not open-source AI if the training code and data information are missing.

What the OSI has said about Meta’s Llama models

The OSI has been explicit about Meta’s Llama models. In a February 2025 post it wrote: “Llama 3.x is still not Open Source by any stretch of the imagination.” Among its reasons, it said the licence “Fails at the Open Source Definition point 5, discriminates against users”.

It said the same about Llama 4 in a LinkedIn post in April 2025: “Llama 4 is still not #opensource”. We have not found an OSI statement on the other licences in the table below, so we do not attribute a verdict to the OSI on them.

What the licences say (September 2026)

Each entry below was read from the model’s official page or its licence file. The summaries are simplified; the exact conditions are in each licence.

Model	Licence	What stands out
DeepSeek-V4-Pro	MIT	The model page says: “This repository and the model weights are licensed under the MIT License”.
OpenAI gpt-oss-120b / 20b	Apache 2.0	The model page describes a “Permissive Apache 2.0 license”. Its usage policy says: “By using OpenAI gpt-oss-120b and gpt-oss-20b, you agree to comply with all applicable law.”
Google Gemma 4	Apache 2.0	Google’s Gemma terms page says “For Gemma 4 terms, see the Gemma 4 license”. Its own Gemma Terms of Use apply to earlier Gemma models, and say you must not use them “for the restricted uses set forth in the Gemma Prohibited Use Policy”.
Meta Llama 4	Llama 4 Community License Agreement	Measured on the Llama 4 release date: if the products or services of a licensee or its affiliates had more than 700 million monthly active users, “you must request a license from Meta”, and you have no rights under the agreement “unless or until Meta otherwise expressly grants you such rights”. Its Acceptable Use Policy also restricts uses including “Military, warfare, nuclear industries or applications, espionage”.
Qwen3.8-2.4T-A95B (Alibaba)	Qwen3.8-Max License	If the licensee or its affiliates run a “Model as a Service or AI Work Assistant business” and their combined revenue exceeds US$50,000,000 in any 12 consecutive months, “the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose”. An AI Work Assistant is defined as “an independent AI-powered product primarily designed for AI-assisted coding or office productivity”, excluding single-purpose tools and assistants built mainly for other domains or as a feature of another kind of product. Internal use is exempt only where it does not make “the Software, its outputs, or its underlying model capabilities available to any third party”.
Mistral Medium 3.5	Modified MIT License	“You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month.”
Kimi K3 (Moonshot AI)	Kimi K3 License	If the licensee or its affiliates operate a “Model as a Service business” and their combined revenue exceeds 20 million US dollars over any 12 consecutive months, “the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose”. Exempt: internal use, defined as use that does not make the Software, “its outputs, or its underlying capabilities available to third parties”, and use through “Moonshot AI’s official products or certified inference partners”.
GLM-5.3 (Z.ai)	GLM-5.3 License	If the licensee or its affiliates operate a “Model as a Service business” and their combined revenue exceeds 10 billion US dollars over any 12 consecutive months, the licensee “must pass Z.AI’s security review before using the Software or its derivative works for any commercial purpose”.

A note on scope for Llama 4: its Acceptable Use Policy also contains an exclusion for “an individual domiciled in, or a company with a principal place of business in, the European Union”, but it applies to the multimodal Llama 4 models, and the policy exempts end users of products built on them. It is not a blanket ban.

Licences can differ within one company

The same company can publish different models under very different terms. On Hugging Face, Mistral Large 3 and Mistral Small 4 are tagged “apache-2.0”, while Mistral Medium 3.5 uses the Modified MIT licence above. Alibaba’s smaller Qwen3.8-27B is tagged “apache-2.0”, and Z.ai’s GLM-5.3-Flash is tagged “mit”, although the larger models in the table use custom licences. Google publishes Gemma 4 under Apache 2.0, while its earlier Gemma models remain under its own terms.

That is why the only reliable check is the licence file for the exact model you are using.

How to read an “open” AI model in two minutes

1. Find the LICENSE file in the model’s official repository, not just the label on the page.
2. Is it a standard licence such as MIT or Apache 2.0, or a custom one with the model’s name in it?
3. Look for thresholds — revenue, monthly users, or types of business that need separate permission.
4. Look for use restrictions — a policy listing things you may not do with the model.
5. Check for training code and data information if “open-source” is being claimed. Weights alone are not enough under the OSI definition.

Our reading: “open” in AI marketing usually means “downloadable”. That is valuable — you can run the model privately, inspect its behaviour and adapt it. But when the word “open-source” is used, it is reasonable to ask for the licence, the training code and the data information, and to judge the claim by what is published.

Sources

• Open Source Initiative, “The Open Source AI Definition – 1.0”; “The Open Source Initiative Announces the Release of the Industry’s First Open Source AI Definition” (28 October 2024); “Open Weights”; “Meta’s LLaMa license is still not Open Source” (February 2025); “Open Weights Are Good. Open Source Is Better.” (16 September 2026).
• DeepSeek, DeepSeek-V4-Pro-0813 model page, Hugging Face.
• OpenAI, gpt-oss-120b model page, Hugging Face; gpt-oss LICENSE and USAGE_POLICY, GitHub.
• Google, Gemma Terms of Use and Gemma 4 licence, ai.google.dev.
• Meta, Llama 4 LICENSE and USE_POLICY, meta-llama/llama-models, GitHub.
• Alibaba Qwen, Qwen3.8-2.4T-A95B LICENSE and Qwen3.8-27B model page, Hugging Face.
• Mistral AI, Mistral-Medium-3.5-128B LICENSE; Mistral-Large-3 and Mistral-Small-4 model pages, Hugging Face.
• Moonshot AI, Kimi-K3 LICENSE, Hugging Face.
• Z.ai, GLM-5.3 LICENSE and GLM-5.3-Flash model page, Hugging Face.

Checked September 2026.

Related: What is a frontier AI model? · AI benchmarks explained · How to spot a fake quote or an invented source