AI 4 November 2026 5 MIN READ
Open Source vs Open Weights in AI
What 'open' really means for AI models: the Open Source AI Definition, open weights, restrictive licences and why the difference changes what you may do.
THE SHORT ANSWER
Open weights means a trained model's parameters are published so you can run and fine-tune it. Open source AI, as the Open Source Initiative defines it, also requires training code and detailed data information. Licences can still restrict use, so read the terms before relying on a model.
On this page
Three things people call “open”
When a company publishes an AI model, “open” can mean three different things, and the difference decides what you are allowed to do.
- Open weights. The Open Source Initiative (OSI) describes open weights as the final weights and biases of a trained network. With them you can fine-tune or run the model. They do not include the training code or the training dataset, and they say nothing about how transparent the data was. There is no single standards-body definition of the term, and researchers at NVIDIA and Stanford use it in the same sense.
- Open licence. The legal terms attached to the weights. A company can publish a model’s weights and still limit who may use it and for what.
- Open source AI. A specific definition, published by the OSI, with specific requirements.
What the OSI definition requires
The OSI announced version 1.0 of its Open Source AI Definition (OSAID) on 28 October 2024 at the All Things Open conference, after a year-long process. It says an open source AI system grants four freedoms: to use, study, modify and share. The precondition is access to the preferred form for making modifications, and that form has three parts.
- Data information: detailed enough information about the training data for a skilled person to build a substantially equivalent system. The dataset itself does not have to be released.
- Code: the code used to train and run the system.
- Parameters: the weights or other settings, under OSI-approved terms.
The definition adds that open source weights must come with the data information and the code used to derive them. Weights alone do not qualify. Stefano Maffulli, who led the process as the OSI’s executive director, summarises it this way: weights and code under an open licence are required, plus enough disclosure about training, but not necessarily the data itself. The definition deliberately does not say which legal mechanism makes parameters free, noting this will become clearer once the legal system has dealt with open source AI.
Licences that limit use
Two of the best-known model families publish weights with restrictions.
- Meta’s Llama 3.1 Community License says that anyone whose products had more than 700 million monthly active users on the release date must request a licence from Meta, which Meta may grant at its sole discretion. The licence page for the later Llama 4 carries the same clause. The licence also says that if you use Llama materials or outputs to train or improve another model that you distribute, the new model’s name must begin with “Llama”. The OSI said publicly that the Llama 2 licence is not open source, pointing to the Open Source Definition’s bans on discrimination against persons or groups and against fields of endeavour.
- Google’s Gemma Terms of Use forbid the uses listed in a prohibited-use policy and require you to pass those limits on to anyone you redistribute to. The terms also let Google restrict use of Gemma “remotely or otherwise” if it reasonably believes the terms are being broken.
Licences change. The Gemma terms page, last modified on 1 April 2026, refers the newest model, Gemma 4, to a separate page, and that page contains the standard Apache 2.0 text. We did not find a Google announcement explaining the change, so check the current licence before relying on it.
Permissive licence, partial openness
A permissive licence is necessary but not sufficient. DeepSeek-R1 publishes its code and weights under the MIT licence, which allows commercial use, modification and distillation. We did not check whether DeepSeek discloses training data or training code to the OSI’s standard, so it would be wrong to call R1 open source AI on the strength of the licence alone.
At the other extreme, the Allen Institute for AI’s OLMo, released in 2024, was a deliberate attempt to be fully open: it released training data and training and evaluation code alongside the weights, unlike most earlier efforts. The OSI’s FAQ lists models it validated against its definition, among them Pythia, OLMo, Amber, CrystalCoder and T5, and others that did not meet it, including Llama 2, Grok, Phi-2 and Mixtral. The OSI stresses that this list is not a certification.
Why the difference matters
The OSI’s view is practical. An open licence lets developers decide where and how to use the technology without asking another party, which it calls sovereignty over the technology they use.
Researchers want something else. James Landay of the Stanford Institute for Human-Centered AI argued in August 2026 that open weights let you run a model locally but not see how it was built or what it was trained on: open distribution, not an open model.
A definition people still argue about
The definition is contested. According to a column in The Register, Bruce Perens, author of the original Open Source Definition, denounced the OSAID in 2024, and Bradley Kuhn of the Software Freedom Conservancy and Richard Fontana of Red Hat have called for it to be repealed, arguing that it created a rift in the free software community. Maffulli answers that the licence on the model matters more than whether the training data is open, and that regulators and adopters act on the licence first and on the builder’s credibility second.
What to watch next
- Read the licence, not the label. Look for user thresholds, prohibited uses and naming rules.
- Data information: whether a release explains what the model was trained on, even when the data is private.
- Who validates: whether the OSI list grows, and whether regulators adopt the definition.
- Licence changes: the same model family can switch terms between versions.
Our take: “open weights” is a distribution choice, not a freedom. It is useful and often generous. But the freedoms to study and to modify depend on the licence and on what else is released with the weights.
Questions and answers
What does open weights mean?
The final weights of a trained model are published, so others can run or fine-tune it. It does not include training code, the training data or any promise about how the data was gathered.
What does the Open Source AI Definition require?
Version 1.0, announced on 28 October 2024, requires access to data information, code and parameters under OSI-approved terms, which together give users the freedom to use, study, modify and share the system.
Is Llama open source?
Meta publishes Llama weights under a community licence with conditions, including a clause for products with more than 700 million monthly users. The OSI said the Llama 2 licence is not open source.
Does a permissive licence make a model open source AI?
Not by itself. A licence such as MIT on the weights is necessary, but the OSI definition also asks for data information and code.
Do you agree?
Take it to Chevlume: write your view, add your sources and let others answer. The closed alpha is by invitation.
Keep reading
The Mechanical Turk: A Chess Machine Hoax
A chess machine from 1770 that was really a hidden person. How the trick worked, who it beat, and how its name ended up at Amazon.
AI / DISCOVERYELIZA, Deep Blue and AlphaGo
Three machines that surprised their makers: a 1966 script that only reflected, a chess engine that calculated and a Go player with a famous move 37.
AI / EXPLAINERWhat Is an AI Agent, Really?
An agent is a model that chooses its own steps and uses tools in a loop. How that differs from a workflow, how it is tested and how it fails.
Sources
- The Open Source Initiative Announces the Release of the Industry's First Open Source AI Definition · Open Source Initiative · Accessed
- The Open Source AI Definition – 1.0 · Open Source Initiative · Accessed
- The license is all you need (for Open Source AI) · Stefano Maffulli (personal blog) · Accessed
- Open Weights: not quite what you've been told · Open Source Initiative · Accessed
- Llama 3.1 Community License Agreement · Meta · Accessed
- Meta's LLaMa license is not Open Source · Open Source Initiative · Accessed
- Gemma Terms of Use · Google · Accessed
- deepseek-ai/DeepSeek-R1 model card · DeepSeek (via Hugging Face) · Accessed
- OLMo: Accelerating the Science of Language Models · arXiv (Groeneveld et al., Allen Institute for AI) · Accessed
- Open Source AI Definition FAQ · Open Source Initiative · Accessed
- Open-Weight Models Aren't Enough. We Need Truly Open Source AI Models for Science and Society. · Stanford HAI · Accessed
- Open weights are not open source: Why AI's favorite label is under dispute · The Register (column) · Accessed
Published
Discuss on X