Skip to content
Read everything about Raspberry Pi 5 and other single board Pi PCs, Pi microcontrollers and other news and updates about Pi accessories.
AI Article · 7 min read

Open Weights Give You the File, and the Licence Decides the Rest

Open weights are a download. Open source in the OSI sense also needs the code and information about the training data. This explainer weighs the OSI definition, Olmo, the gpt-oss and Llama licences, Epoch's four-month lag and what running a...

Laptop terminal with a command being typed beside a small desktop PC with a glowing graphics card

It has been more than a year since OpenAI released gpt-oss and called its two models open-weight, and the word open is still attached to model releases that contain very different things. I think the word deserves more suspicion than it gets, and that the useful question when a lab calls a model open is a plain one about what actually lands on your disk.

Almost always it is the weights: the trained numbers that make the model work, packaged so anyone can download them and run the model on their own hardware. That is what open weight means. A download of that kind falls short of open source as the Open Source Initiative defines it for AI, which also asks for the code that produced those numbers and information about the data they were trained on. I think that gap matters less to most readers than the licence that comes with the file, because the licence is what decides what you are allowed to build.

Is open weight just open source with a trendier name?

The Open Source Initiative wrote its Open Source AI Definition 1.0 to settle what open source should mean for AI, and its bar sits well above a download. A system qualifies only if it grants four freedoms: to use, study, modify and share it. The first is worded with care, as the right to use the system "for any purpose and without having to ask for permission." So that studying and modifying are possible in practice, the definition asks for three components together. They are sufficiently detailed information about the data used to train the system, the complete source code used to train and run it, and the model parameters, such as weights or other configuration settings.

Hold an open-weight release up against that and the weights meet one requirement out of three. You get the parameters, and you can use the model and fine-tune it on your own material. The weights alone do not let you rebuild the model or check what went into it, because that takes the training code and information about the data, which the OSI lists as separate requirements. The popular definitions are vague on exactly this point. Stanford HAI's glossary entry says the core components are publicly released so that anyone can download the model, and it does not say whether data or code are among them.

Releases that go much further do exist, which is why I don't treat the distinction as pedantry. The Allen Institute for AI's Olmo publishes the weights, code and training data mixture, along with reports and checkpoints, and the institute calls that data "the fully open mixture used to train Olmo from scratch." I grant that most people who download a model will never train one from nothing, but anyone who wants to study why a model answers the way it does, or whether a test set slipped into its training data, needs to know what that data was, and a weights-only release does not tell them.

Can I really do anything I like with the download?

That depends on which download it is. OpenAI released gpt-oss on 5 August 2025 in two sizes, gpt-oss-120b and gpt-oss-20b, and calls them open-weight language models. Both come under Apache 2.0, a permissive licence that allows commercial use, modification and redistribution as long as you keep the licence text and notices.

Meta's Llama 3.1 licence reads very differently. A company whose products had more than 700 million monthly active users in the preceding calendar month has to ask Meta for a licence. No hobbyist will come near that number. Still, it means your right to use the model depends on who you are, and I think that sits badly with the OSI's wording, which promises use "without having to ask for permission." The same licence also requires you to prominently display "Built with Llama" on a related website, user interface or blog post, so anything built on it carries Meta's name.

Every model page on Hugging Face shows its licence next to the files, and I think that one line deserves more attention than the benchmark scores above it. A permissive-sounding name on the label does not settle the question. RedMonk has described one model as "not open source, MIT license branding notwithstanding," and points out that a single vendor's model family can carry a mix of licences. You have to read the licence for each model, because the brand tells you nothing.

RedMonk makes a harder point as well. It is not settled law that copyright can be applied to weights at all, which means licences written for source code may not fit them cleanly. I don't read that as a loophole. For anyone shipping a product, the sensible course is to follow the licence as written and not bet on a court reading it narrowly.

Is a closed API still the sensible default?

A closed API sends your prompts to a vendor's servers, and the model never leaves them. Its strongest argument is that it gets the newest models first. Epoch AI has measured how far the open side trails: in 2026, open-weight models lagged frontier closed models by an average of four months, or 8 ECI points. On some skills the gap is longer, and Epoch puts open models six to seven months behind on cyber capabilities.

Four months does matter if your work depends on the strongest reasoning available today, and I grant that an API also spares you the hardware, the updates and the electricity. But another of Epoch's findings changes how I weigh that lag. Epoch reports that frontier AI performance becomes accessible on consumer hardware within a year.

The trade I care about more is privacy. Prompts sent to an API leave your machine and go to someone else's servers under their terms, and you pay per use at their prices. With the weights on your own disk, nothing leaves the machine, and your running cost is hardware you control. I think that outweighs a four-month lag.

What does running it yourself demand?

For gpt-oss the hardware question has a clear answer, because OpenAI published the figures with the release: gpt-oss-120b runs within 80GB of memory, and gpt-oss-20b needs only 16GB. The larger model needs more memory than any ordinary home graphics card has. The smaller one is within reach of a well-specified desktop or laptop. On a Raspberry Pi as on a PC, OpenAI's 16GB figure for gpt-oss-20b is the number to set against the memory you have.

The software is the easy part. Ollama is an MIT-licensed runner built on llama.cpp, so the tool itself comes with no licence strings of its own, and it can load models in the GGUF file format straight from Hugging Face.

The Hugging Face documentation gives the command below, with the account and repository names from the model page taking the place of the braces.

ollama run hf.co/{username}/{repository}

That one line hides a decision. By default, Ollama uses the Q4_K_M quantisation scheme when the repository includes it. Quantisation stores each weight at lower precision, roughly four bits in this case, which shrinks both the file and the memory it needs at some cost in accuracy. Quantised files like these are how many models get small enough for a home machine.

So which one should you pick?

If you want the newest capability and no hardware to look after, a closed API is the reasonable choice. If your prompts cannot leave the building, or you want to fine-tune a model on your own material and ship it, an open-weight release fits better, and the licence on the model page decides what you may do with it.

Then there is the smaller group who need to understand a model rather than just run it, to audit what it learned or to reproduce it. For them, only a fully open release will do, because only that kind of release includes the training code and information about the data, as Olmo does with its published training mixture. It is a minority of readers, but it is the group the word "open" was originally meant to serve, and I would like to see labs stop using that word for releases that leave them out.

Choosing open weights means giving up the newest frontier model and a vendor's hosting, and in return you get control over the model, privacy for your data and a licence you have read for yourself.

Written by Barış Aydın

I write about artificial intelligence, mostly whether each week's big announcement deserves the attention.

1 article →