Informopedia Explains: this week, a French company gave away something its rivals charge a fortune to rent — the inner workings of one of the world’s most powerful AI models. Mistral’s release of Large 4, a model with roughly a trillion parameters, is being called “open.” That word is doing a lot of work. Here is what it actually means.
What a “weight” is
An AI model is, at its core, a vast web of numbers — the “weights” — tuned by training on enormous amounts of text, code and images. Given a prompt, the model crunches those numbers and predicts what should come next. Whoever holds the weights holds the model: they can run it, study it, adapt it and build on it without asking anyone’s permission.
Most frontier models are closed. You can use them through an API, but you cannot see or download the weights, you cannot run the model on your own computers, and your data passes through someone else’s servers. Open-weight models flip that: the weights are published, usually on a platform like Hugging Face, and anyone with enough computing power can download and run them.
Open-weight is not open-source
This is the distinction that matters. True open-source software comes with its source code, its training data and a licence that lets anyone do anything. Open-weight models usually come with only the weights — not the training data, not the full training recipe — and a licence with conditions. Mistral’s licence, for example, lets anyone use, modify and commercialise Large 4, except businesses that offer the model itself as a hosted service, which need a separate commercial licence.
That is not a trick; it is a business model. Mistral gives away the model to build an ecosystem, and charges the companies that would resell the model itself. But readers should not confuse the two: “open-weight” means you can run it, not that you can fully reproduce it.
Why a trillion parameters, and why so few active?
Large 4 has about a trillion parameters in total but only around 52 billion active for each token — each word it generates. This is the mixture-of-experts design: the model is divided into many specialist sub-networks (“experts”), and only a few are consulted for each word. It is like a hospital with a thousand specialists where only four enter the room per patient. The result is a model that can be both enormous in knowledge and comparatively affordable to run — which is exactly what makes it practical for companies to self-host.
Why anyone would give this away
Three reasons. First, sovereignty: a government, hospital or bank that cannot send sensitive data to a foreign cloud can run an open-weight model inside its own data centre. Second, customisation: with the weights in hand, developers can fine-tune the model on their own data — legal documents, medical notes, a company’s codebase — in ways an API never allows. Third, trust: researchers can probe the model for biases and flaws instead of taking the vendor’s word for it.
The catch
Capability cuts both ways. Mistral itself says Large 4 has “a potential for uplift to conduct real attacks” in cybersecurity, and it is keeping a more aggressive penetration-testing version restricted to vetted partners. The public model is pitched at defence — finding and patching vulnerabilities — but the same skills can be turned around. Every open-weight release is therefore a judgement call: the benefits of wide access against the risk of misuse by people who would never have been allowed near a closed model.
Why this matters
Mistral’s release will not make closed models disappear; the biggest labs still keep their best systems behind APIs. But it raises the floor for everyone else. A university in a country without a frontier AI lab, a startup that cannot afford API bills at scale, a government that wants AI without dependence — all of them just got a much stronger starting point. The open-weight race is, in the end, a race over who gets to build the future: a handful of companies, or anyone with a server rack and an idea.
