What is open-source AI and how is it different?

Open-source AI means the underlying model, and often its training recipe, are published for anyone to inspect, copy, and modify. The opposite is a closed or proprietary model, where you can only access the AI through an API or app and the inner workings stay secret. Think of it like a recipe book versus a restaurant that will only sell you the finished dish.

What exactly gets "opened up" in an open-source AI model?

At minimum, the model weights are released publicly. Weights are the millions or billions of numerical values the model learned during training, the numbers that actually make it work. Some projects also release the training code, the dataset, and documentation.

Not every project releases all of those things at once. Meta's Llama models, for example, release the weights and allow broad use, but the full training data is not published. Meta's own research page for Llama describes this in detail. True open-source purists would argue that releasing weights alone is closer to "open weights" than full open-source, but in everyday conversation the terms are used interchangeably.

How is open-source AI different from closed AI like GPT-4?

With a closed model, you send your text to a company's server, it does the thinking, and sends an answer back. You never see the model itself. Open models let you download the whole thing and run it on your own computer or cloud server.

That single difference has big knock-on effects. You can fine-tune (specialise) the model on your own data, audit it for bias or errors, and keep your data entirely on-premise, meaning it never leaves your building. A hospital, for instance, could run an open model locally so patient records never touch an outside server.

What are the real benefits?

Four things stand out. First, cost: once you have the model, inference (running it) can be far cheaper than paying per-query API fees. Second, customisation: you can reshape the model's behaviour in ways a commercial API does not allow. Third, transparency: researchers can study how the model reaches its answers, which matters enormously for safety research. Anthropic's alignment research highlights why being able to inspect model internals is valuable. Fourth, no vendor lock-in: you are not dependent on one company staying solvent or keeping prices stable.

What are the downsides and risks?

Open models can be misused more easily, because anyone can remove safety guardrails that a company would normally enforce. They also require technical skill to run well, and you are responsible for keeping them updated and secure. Smaller organisations may find the infrastructure burden heavy.

NIST's AI Risk Management Framework notes that transparency cuts both ways: it helps defenders understand a system, but it also helps attackers understand how to abuse it.

Open-source vs closed AI: a quick comparison

Feature	Open-source / open weights	Closed / proprietary
See the model internals	Yes	No
Run it on your own hardware	Yes	No
Customise and fine-tune	Yes, freely	Limited or not at all
Data stays on your servers	Yes	Usually no
Upfront technical effort	High	Low
Typical cost model	Infrastructure costs	Pay-per-query fees

A real-life example of open-source AI in action

A legal tech startup wants to build a document-review assistant. Using a closed API means every contract it processes passes through someone else's servers, which raises client confidentiality problems. Instead, the team downloads an open-weights model, fine-tunes it on legal language, and runs it inside their own secure environment. Clients never need to know, and no data leaves the firm's control.

What does the licence actually allow?

This is where things get nuanced. "Open-source" does not always mean "do anything you want". Some models carry licences that restrict commercial use or require you to open-source any modifications you make. The Open Source Initiative's definition requires, among other things, that the licence not restrict anyone from selling or giving away the software. Many popular AI model licences fall short of that bar, which is why the phrase "open weights" is becoming more precise.

Always read the licence before building a product on top of a model.

Privacy catch, in plain words

Running an open model locally is actually the most private option available, because your prompts never leave your machine. The catch is that downloading large model files (often 10 to 70 gigabytes) and running them requires decent hardware. Consumer laptops can handle smaller models; large ones need a powerful GPU or a rented cloud server.

Common questions

Is open-source AI free to use?

Downloading the model is usually free, but you still pay for the electricity and hardware (or cloud compute) needed to run it. Commercial licences on some models may also require a fee for business use.

Can I trust an open-source AI to be safe?

Open models can be audited by anyone, which helps catch problems, but they can also have safety guardrails removed by anyone. Safety depends heavily on how carefully you deploy and configure the model for your specific use case.