What Open Weights Does Not Mean

“Open weights” has become shorthand for trustworthy, free, and yours. It is a genuinely important property and it is worth understanding precisely, because the gap between what it guarantees and what people assume it guarantees is where the unpleasant surprises live.

Here is what the term actually covers, and four things it does not.

What It Does Mean

Open weights means the trained parameters of the model are published for download. You can obtain the file, run it on your own hardware, inspect its outputs, and keep using it after the vendor loses interest or changes their pricing.

That is a real and valuable set of properties. It is the difference between a capability you have and a subscription you rent, and it is why open weights matters for anyone with a confidentiality obligation. Why open weights matter for workplace privacy makes that case.

Everything below is what the term does not carry with it.

It Does Not Mean Open Source

This is the most common confusion and the most consequential.

Open source, as defined by the Open Source Initiative, requires among other things that the licence not discriminate against fields of endeavour or persons, and that redistribution be permitted without additional restrictions. Many popular open weights licences do not meet that bar.

Some impose acceptable use policies restricting what you may generate. Some restrict commercial use above a user threshold. Some require attribution or naming conventions in derived models. These are licence terms, they are enforceable, and they are not open source in the OSI sense even when the download is free.

The models that genuinely are permissively licensed say so plainly: Apache 2.0 or MIT. Wrivio’s local tiers are Apache 2.0 Qwen3 models, which is why the blog names them openly. An earlier version of the product shipped a model under a research licence that prohibited commercial use, which was the wrong choice and was replaced.

If commercial use matters to you, read the licence rather than the announcement. Open weights model licences compared covers the common ones.

It Does Not Mean The Training Data Is Available

Weights are the output of training, not the input. Almost no major open weights release includes the dataset, and many do not describe it in enough detail to reconstruct.

This matters for two reasons. You cannot audit what the model learned from, so questions about copyright, bias, or contamination cannot be answered by inspection. And you cannot reproduce the model, which means “open” here does not carry the scientific reproducibility that the word implies in other contexts.

The term of art for a release that includes data and training code as well as weights is “open source model”, and it is rare. Assume you are getting weights only unless told otherwise.

It Does Not Automatically Mean Private

An open weights model can be run privately. It is not private by virtue of being open.

If you use an open weights model through a hosted API, your text goes to that host exactly as it would with a proprietary model. The licence of the weights says nothing about the data handling of the service in front of them.

Privacy is a property of where the model runs, not of how it is licensed. A proprietary model running entirely on your machine would be more private than an Apache-licensed model running on someone else’s server. The two properties are independent and get conflated constantly.

This is why the useful question is never “is it open weights” but “which machine processes my text”. How to tell if an AI tool really runs offline is the practical version of that question.

It Does Not Mean You Can Run It

A published model you cannot execute is a curiosity. Plenty of open weights releases are far beyond consumer hardware: a model with hundreds of billions of parameters is open in the sense that you could download it and impractical in the sense that you cannot.

Parameter count, quantization, and memory requirements decide whether an open model is available to you in practice. A 4B model you can run beats a 200B model you cannot, and the marketing rarely makes that distinction.

How to choose an open weights model for your laptop is the sizing exercise, and can you run a trillion parameter model locally is the honest answer to the ambitious version of the question.

What To Actually Check Before You Rely On One

Four questions, in order, each answerable in a minute from the model card.

What licence, by name, and does it permit your intended use commercially? Apache 2.0 and MIT are unambiguous; anything custom needs reading.

What are the memory requirements at the quantization you intend to use, and does your hardware meet them with headroom?

Where will it actually run, and does that satisfy the confidentiality obligation you are trying to meet?

And is there a model card at all? A release with no documentation of intended use and limitations is telling you something. How to read an AI model card covers what a good one contains.

Common Questions

Is open weights the same as open source?

No. Open weights means the trained parameters are downloadable; open source is a licence property with specific requirements. Many open weights licences restrict use in ways that would not qualify as open source.

Does open weights mean the training data is public?

Almost never. Weights are the output of training, and most releases publish neither the dataset nor enough detail to reconstruct it, so the model cannot be independently reproduced or audited.

Is an open weights model automatically private?

No. Privacy depends on where the model runs. Using an open weights model through a hosted API sends your text to that host exactly as a proprietary model would.

Which open weights licences are safest for commercial work?

Apache 2.0 and MIT are permissive and unambiguous. Custom licences from major labs often add acceptable use restrictions or commercial thresholds, so read them rather than assuming.

Download Wrivio for Windows to run Apache 2.0 licensed models entirely on your own machine, where the licence and the privacy property both hold.