What is open-source AI?

Two shops sell the same bread, and the first sells you a loaf. The second sells you a loaf and hands you the recipe, the flour list and the oven settings. You can bake it at home, and you can add more salt. You can open your own bakery down the road.

The bread is the same on the day you buy it, but everything after that day is different.

Open-source AI applies that idea to models. Somebody builds a model, then publishes enough of it for you to run it, study it and change it yourself. Most of the models people argue about are large language models.

What gets published, and what usually does not

An AI model has several parts, and each part can be shared or held back.

• The weights. These are the numbers the model learned during training, and they are the model itself.
• The code that runs the weights, and the code that trained them.
• The training data, or at least a list of where it came from.
• The licence. This is the legal note that says what you may do with all of the above.

Most models called open today publish the weights and a licence. Fewer publish the training code, and very few publish the training data. Why does the data stay hidden? Because it is huge, it is costly, and it often holds material the publisher would rather not list in public.

Open weights and open source are not the same thing

This is the argument at the centre of the subject, and it is worth being exact.

Open source has a settled meaning in software. You get the source, and you may run it for any purpose. You may change it, and you may pass it on. Anyone may do this, including a rival firm.

However, many AI models fall short of that test. You can download the weights, yet the licence blocks some uses. Or the recipe is missing, so you could never rebuild the model yourself even with a room full of machines.

A fairer name for those is open weights. Therefore read the licence before you use the word. OPEN WEIGHTS AND OPEN SOURCE ARE DIFFERENT THINGS.

How you actually run one

You need three things, and they are the weight file, a program that can run it, and a machine with enough memory.

The weight file is a download and the program is free. However, the machine is the part that costs money. A large model wants a graphics card with plenty of memory, or several of them working together.

There is a trick that helps, called quantisation. It stores each number with less detail, the way a photo saved at lower quality takes up less space. Therefore a quantised model uses far less memory, and it answers a little less well. Many small models run on an ordinary laptop this way.

What open models are good at

• Privacy. Your data never leaves your machine, because the model comes to the data.
• Cost at scale. You pay for hardware once instead of paying per question forever.
• Control. Nobody can change the model under you or switch it off next year.
• Study. Researchers can look inside and test claims. Closed models do not allow that.
• Adaptation. You can train the model further on your own examples.

That last point matters most for languages and subjects the big labs pay little attention to. If your language is served badly, you can improve the model yourself, but with a closed model you can only ask and wait.

Open source has been the normal way to build machine learning tools for years. The tools that train models, the libraries that move the numbers, and the operating system on most servers are all open. Only the models themselves are in dispute.

What open models are bad at

The very best models are usually closed, at least for a while, and open models tend to follow behind. The gap has narrowed a great deal, and it moves every few months, so check for yourself rather than trusting a claim from last year.

You also take on work. Nobody patches the model for you, and nobody filters harmful output for you. Nobody keeps the server running at three in the morning. Free to download is not free to run.

And openness cuts both ways, because anyone who holds the weights can strip the safety training out of them. That is a real cost of the approach, and it is honest to say so plainly.

Where you already meet open models

More often than you think. There is the speech recognition that types what you say, and the small model that guesses your next word. There is the blur behind you in a video call, and the tool that pulls text out of a photograph of a receipt.

Many of these run on published models that somebody adapted. You never see a name, because the model sits inside a product. Read our piece on neural networks for what is happening under all of them.

Why companies give models away

It looks odd, with years of work handed out for nothing. There are three plain reasons.

The first is that a free model can weaken a rival who sells one. If the thing your competitor charges for becomes free, their business gets harder.

The second is standards. Whoever gives away the tool everybody builds on gets to shape how the field works, and gets the best engineers applying for jobs.

The third is improvement. Thousands of outsiders test the model, find its faults and publish fixes, and that is unpaid work of very high quality.

How to check a model before you use it

1. Read the licence. Look for limits on business use, on user numbers, and on training other models with it.
2. Look for a model card. That is a short document saying what the model was made for and where it fails.
3. Check the file size against your hardware before you download tens of gigabytes.
4. Test it on your own real questions, not on the examples in the announcement.

Then run one small model on your own machine this week. Nothing teaches the trade-offs faster. You will feel the speed, the memory use and the quality in a single afternoon, and you will stop guessing.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.