Open-Weight Is Not Open Source: What You Actually Download

There are two terms used interchangeably across the AI industry, and the interchange produces the wrong expectations. Open-weight means the model's weights can be downloaded. Open source means considerably more than that, and almost no large AI model comes close.

The difference is not just vocabulary. That difference decides whether you can audit the model you are using, whether that means its training data, its filters, or its biases. And because AI models are now a base layer under many products, the right to audit is not a technical detail anyone can skip.

Two words that get swapped for each other

Open-weight is, to be fair, a very clear term. The weights exist, you download them, you run them. The term is used by the Open Source Initiative itself, with a definition that refers to the final weights and biases of a trained neural network. Once you have the weights, you can change the model's output, adapt it to your own data, or run it on your own machine without someone else's service.

Open source has a much heavier definition, and that definition is not new. The OSI has defined four freedoms since 1998, and they still apply. You may use it for any purpose. You may study how it works. You may modify it. You may redistribute it, with or without changes.

The problem, and this is the core of the article, is that almost nothing published under the open source label actually qualifies. Writing in The New Stack in October 2024, one observer noted that the open-wash releases from companies like OpenAI and Meta do not make the OSAID grade. The OSI itself says that open source models and open weights must include the data information and the code used to derive those parameters.

OSAID 1.0: the strictest definition available

Open Source AI Definition version 1.0 was released on 28 October 2024 at All Things Open 2024 in Raleigh, North Carolina. The process took roughly two years, with the last year spent on co-design involving the global community.

That definition requires three things at once. First, data information detailed enough that a skilled person could build a substantially equivalent system. The detail requested includes the provenance of the data, its scope and characteristics, how it was obtained and selected, the labelling procedures, and how data was processed and filtered. Second, the complete code used to train and run the system, including data processing code, training code with the arguments and settings used, validation and testing, supporting libraries such as tokenizers, inference code, and the model architecture. Third, the model parameters themselves, the weights, under OSI-approved terms.

The keywords are in the first two. A free licence alone is not enough, and open weights alone is not enough. Both must come with the information that lets someone rebuild an equivalent system.

The definition was not without controversy. Bruce Perens, author of the original 1998 Open Source Definition, rejected it in 2024 on openwashing grounds. RedMonk analyst Stephen O'Grady framed the core problem this way. Source code is a precisely and narrowly bounded subject area, and AI projects are not.

What you actually receive when you download

Now for the part that gets skipped most, which is what is actually inside the package when someone calls their model open-weight.

Look at the third row from the bottom. The real prize is the training data information. The data itself cannot be downloaded, and almost nobody genuinely wants to publish terabytes of text scraped from the internet. What is realistic, and what OSAID uses, is a summary detailed enough.

That is where the difference lands. Open-weight gives you the model. Open source, if not literally, gives you the ability to judge whether the model can be trusted.

The licences are what actually bind

Downloading is not the only thing that binds. The licence is, and the licences are where the difference shows up most clearly.

The first row is the one most often missed. Llama 3 is a licence with conditions, not a free one. If your product has more than 700 million monthly active users, you must request an additional licence from Meta, and Meta decides whether to grant it. You also have to say Llama at the start of the name of any derivative model, and you have to carry its use policy along.

Compare that with MIT. DeepSeek-R1 shipped under MIT, and its release notes state that its API outputs may be used for distillation. That is a large difference in real use, and the difference is not aesthetic.

The interesting part is that MIT and Apache 2.0 are now used precisely by the companies that do not use the phrase open source. DeepSeek uses MIT. Qwen uses Apache 2.0. Both are open-weight models. Under the definition the industry actually uses, open source, neither would pass. Under open-weight, they are the cleanest entries in the table.

Analysis: why the gap matters rather than being semantics

The most common question is whether this distinction matters in practice or is just definitional argument. The answer depends on who is asking.

For a developer shipping a product, the gap is real. An open-weight model means you have that component, and the component comes with licence terms that actually bind. An open source model means you also have the right to understand that component, which is enormously valuable when the model is taking decisions about money, health, or people's rights.

The third and fifth rows are the heart of the whole argument. Running and fine-tuning are within reach, and that is what makes open-weight models useful. Auditing the training data and checking the alignment method, the two things that matter most when a model decides things about people's lives, are not possible without the information OSAID asks for.

For a researcher writing a paper about model behaviour, the gap decides whether the paper is possible. You can audit the weights, test the behaviour, run the model on your own data. You cannot verify training data claims if the training data was never published.

And for a regulator, the gap has a concrete legal effect. The EU AI Act grants general purpose AI models an exemption from some obligations if they are released under a free and open source licence permitting access, use, modification, and distribution, provided a public summary of training content is published. That exemption depends on openness being real, not on a licence label.

There is one thing worth being honest about. The line is moving. The Register wrote in September 2026 that OSI's treatment of weights leaves a large hole, and the Linux Foundation's Open Model, Data, and Weights licence has been submitted to the OSI with contributors from Amazon, Meta, IBM, Microsoft, and Nvidia. So this definition is not settled.

How to check one yourself in five minutes

If you are evaluating a model, there are four files worth opening, and all of them are free.

First, find the licence file in the model's repository rather than on a marketing page. Read the conditions yourself. If there is a sentence about a user threshold, about the model name, or about a ban on using outputs to train another model, that is information your legal team needs.

Second, look for a folder with training code rather than just inference code. Inference code ships in plenty of distribution packages. Training code, complete with its arguments and settings, is almost never present in an open-weight model.

Third, look for a training data summary. If there is none at all, the model does not claim to be open source, and nobody needs to evaluate it. If there is one, check whether it only lists well-known dataset names or whether it explains how the data was filtered and processed. The second kind is what makes a summary useful.

Fourth, count the derivatives. Alibaba's Qwen family has been reported at around 113,000 derivative models on Hugging Face, and that figure comes from a third party so it cannot be audited, but the direction is clear. A model family with thousands of derivatives and a clean licence usually means the base was free to build on. A family with a handful of derivatives usually means the licence binds.

What happens next

Three things are worth watching.

First, whether OpenMDW clears OSI's licence review. If it does, the industry gets a more detailed definition of openness than it has now, and that definition has substantial industry backing.

Second, whether the open source label stops being used for models that do not qualify. The New Stack has already written that this is what happened. A tighter definition coming back will slow that down without stopping it.

Third, and this is the least discussed, whether already-released weights stay usable. Several modern licences still carry clauses about models trained with the help of other models' outputs, as seen in Llama 2. If such rules bind, you cannot use a newer model to build a successor to an older one, and boundaries drawn years ago still apply in a world that has moved on.

The most commonly skipped part of all this is that nothing changes about the weights. The weights can be downloaded in both cases. What differs is the information you get alongside them, and who is entitled to ask for it.

If you only use a model to argue a case in a meeting, you may not feel the gap. If you are signing contracts that concern other people's rights, or writing analysis that needs evidence, the gap decides whether you have any ground to stand on at all.