Open-weight vs open-source models: the real difference

"Open" is one of the most overloaded words in the model world. A model gets called "open" and people assume it means what "open source" has meant in software for decades: you can see everything, change anything, and reuse it freely. With models, that assumption is usually wrong. Most models described as open are open-weight, which is a narrower and more specific thing than open-source — and the gap between the two terms decides what you are actually allowed and able to do. Treating them as synonyms leads to wrong expectations about transparency, reproducibility, and your legal rights.

This piece untangles the two. It explains what a model is made of, what "open-weight" and "open-source" each really release, why the distinction matters in practice, and how to read the fine print before you commit.

What a model is actually made of

To see the difference, you have to see the ingredients of a model. There are several, and "openness" can apply to any subset of them.

The weights are the learned parameters — the giant collection of numbers the model arrived at through training. They are what you load to actually run the model. The architecture is the design of the network: how it is structured and how data flows through it. The training code is the software used to train it. The training data is the corpus it learned from. And the training recipe is everything else: the configuration, the procedures, the choices that made the difference between a mediocre model and a good one.

The crucial point is that releasing the weights does not release the rest. You can hand someone the finished numbers without handing over the data, the code, or the recipe that produced them. That separability is exactly what the two terms carve up differently.

What "open-weight" means

An open-weight model is one whose weights are made available to download and run. You get the finished model — the numbers — and you can load it, run it on your own hardware, and use it in your products subject to its license. This is genuinely valuable: it means you are not locked into one provider's servers, you can run the model privately, and you can build on it without asking permission for each use.

But open-weight typically stops at the weights. You usually do not get the training data, often do not get the full training code, and rarely get the complete recipe. So you can use the model, but you cannot reproduce it from scratch, you cannot fully audit how it came to be, and you cannot independently verify what went into it. You have the artifact, not the factory that made it. For most practical uses that is enough; for transparency, reproducibility, or deep trust, it is not.

What "open-source" means

"Open-source," used strictly, sets a higher bar inherited from software. The established meaning of open source is not merely "you can look at it" — it is a set of freedoms: to use the thing for any purpose, to study how it works, to modify it, and to redistribute it, including modified versions, without restrictive conditions. Applied honestly to a model, that implies releasing enough of the ingredients that someone else could genuinely study, rebuild, and adapt it — not just run the finished weights.

This is a meaningfully stronger claim than open-weight, and it is rarer. Many things marketed as "open-source models" release only the weights under a license that restricts how you may use them, which fails the stricter definition on two counts: incomplete ingredients and restrictive terms. The word carries the prestige of decades of open-source software, which is precisely why it gets applied loosely to models that do not earn it.

Why the distinction is not pedantic

This matters because the two terms create different rights and different abilities, and conflating them sets you up for three concrete problems.

First, legal rights. A model's license governs what you may actually do — whether you can use it commercially, build a competing product, redistribute it, or use it above a certain scale. "Open" in the marketing does not tell you any of this; the license does. A model can be freely downloadable and still come with restrictions that rule out your intended use.

Second, transparency and trust. If you need to know what data a model learned from — for compliance, for bias review, for understanding failure modes — open-weight alone will not give it to you. Only fuller disclosure of data and recipe supports that kind of scrutiny. Assuming "open" means "inspectable all the way down" leaves you unable to answer questions you may be obligated to answer.

Third, reproducibility. If your work requires being able to retrain, verify, or build the model again from its ingredients — common in research and in regulated settings — weights alone are not enough. You need the code, data, and recipe. Expecting to reproduce an open-weight model and discovering you cannot is a costly surprise to have late.

How to read the fine print

Because the labels are unreliable, judge a model by its actual terms rather than its adjective. A few questions cut through the marketing.

Ask what was actually released. Just the weights? Also the training code? Any information about the data? The architecture details? The answer places the model on the spectrum far better than the word "open" does.

Ask what the license permits and forbids. Read it for commercial use, redistribution, modification, scale limits, and any field-of-use restrictions. This is the part with legal force, and it is where loosely-labeled "open" models most often disappoint.

Ask what you actually need. If you only need to run a capable model privately and ship a product, open-weight under a permissive-enough license may be exactly right and the deeper questions are moot. If you need to audit, reproduce, or freely redistribute, you need more, and you must confirm you have it rather than assume.

A spectrum, not two boxes

It helps to drop the binary entirely. Openness is a spectrum running from a model you can only call through someone else's API, through a downloadable open-weight model under a restrictive license, to a permissively-licensed open-weight model, to a model that also discloses code and data, all the way to one that releases the full ingredient list under genuinely free terms. Each step grants more freedom and more transparency. The single word "open" collapses this whole range into one syllable, which is exactly why it misleads. Locate a model on the spectrum and you will know what you are getting.

The takeaway

Open-weight and open-source are not synonyms. Open-weight means you can download and run the finished weights — useful, but it usually stops there, leaving the data, code, and recipe behind. Open-source, used properly, means a fuller set of freedoms and ingredients, and it is rarer than the marketing suggests. The distinction governs your legal rights, your ability to inspect, and your ability to reproduce. Ignore the adjective, read the license and the release notes, and match what was actually provided against what your project genuinely needs. The word "open" is a starting question, never the answer.