What is Mistral AI?

Mistral AI is a French artificial intelligence company, based in Paris, known for its open language models. On 10 October 2023 it published "Mistral 7B," the paper introducing its first model, which closes with a sentence more important than everything else in it: "Our models are released under the Apache 2.0 license."

That sentence is worth pausing on, because it is rare. Apache 2.0 is a genuinely open-source licence, one of those approved by the Open Source Initiative: no user thresholds, no restrictions on what you may use it for, no discretionary permissions. Most models advertised as open do not reach that bar.

What Mistral 7B is, with the exact quantifiers

It is a 7-billion-parameter model "engineered for superior performance and efficiency." And here the abstract deserves precise quotation, because it contains two distinct claims that routinely get merged into one:

"Mistral 7B outperforms Llama 2 13B across all evaluated benchmarks, and Llama 1 34B in reasoning, mathematics, and code generation."

Against Llama 2 13B: on all of them. Against Llama 1 34B: in three specific areas, not generally. Two different scopes, and the paper separates them carefully.

Notice what happens if you summarise both as "beats several models on several benchmarks." It sounds cautious, and it is false twice over: it undersells the first and oversells the second. Flattening two precise claims into one vague one is not stopping halfway — it is erring high and low simultaneously, while looking prudent. It is one of the commonest ways to falsify without lying, and it is why the quantifiers in a technical document must be copied, not paraphrased.

The underlying point is that a 7-billion-parameter model competes with others of 13 and 34 billion. In this field size is paid for in memory, in electricity and in price per query, so punching above your weight is not a sporting curiosity: it decides whether the model fits on ordinary hardware.

The two techniques that make it cheap

The paper attributes that efficiency to two specific mechanisms, and they are worth understanding because they now appear in almost every new model.

Grouped-query attention exists, in the abstract's own words, "for faster inference." The idea: rather than each attention head keeping its own copy of the information it consults, several share one. Less is stored, less is read, answers come sooner.

Sliding window attention allows the model "to effectively handle sequences of arbitrary length with a reduced inference cost." Instead of every word attending to all preceding ones — which gets expensive fast as text grows — each attends only to a window of the nearest, and information travels further by hopping from layer to layer.

Both are engineering decisions, not scale. That is the message worth keeping: part of what today separates a good model from an expensive one is not how much data it swallowed but how it is built inside.

The paper also mentions an instruction-tuned variant, Mistral 7B — Instruct, which "surpasses the Llama 2 13B — Chat model both on human and automated benchmarks."

Apache 2.0 against Llama's licence: two cases side by side

Here is what you actually take away, and seeing it requires placing two documents next to each other.

The Llama 2 licence, which we analysed in detail in another piece, contains two clauses the Open Source Definition does not allow: a threshold of 700 million monthly active users beyond which you need permission Meta grants "in its sole discretion," and a ban on using the model or its outputs to improve any other language model. Those are open weights under a bespoke licence.

Apache 2.0 has none of that. It is a standard text, written in 2004, that any lawyer recognises, that discriminates by neither size nor sector, and that expressly grants patent rights as well. The practical difference for anyone building on top is enormous: with Apache 2.0 there is no future condition lying in wait for the day your product grows.

What you can actually do, and what you take on

This is worth bringing down to ground level, because "open" sounds like a virtue and is really a division of labour.

What you gain from a model whose weights ship under a standard licence: you can run it on your own hardware, which means the data you feed it never leaves your network — a far sturdier answer to the privacy question than any contractual clause. You can tune it to your domain. You can pin a version and stay on it as long as you like, rather than discovering one Monday that the model behind your product has changed its behaviour. And you can leave for another with nothing to renegotiate.

What you take on in exchange, and this is rarely spelled out: the whole responsibility becomes yours. There is nobody to complain to if the model produces something harmful, no team shipping security patches, no vendor answering to a regulator on your behalf. Infrastructure, compute cost and behavioural monitoring all fall to whoever deploys it.

Which is why "open" is not automatically better: it is a different division. For an organisation with a technical team and sensitive data it almost always pays; for one that just wants a feature that works, almost never. The useful question is not which of the two models is better, but which of the two burdens you can carry.

The capability you take away

And now the part that outlives this model and this company:

"Open" is not a property of the company but of each specific release.

The same Mistral that published its first model under Apache 2.0 later shipped models under bespoke, restrictive licences. That is neither betrayal nor incoherence: it is an ordinary commercial decision, taken separately for each product. What is wrong is what we readers do — filing companies under "the open ones" or "the closed ones" and then assuming we know the licence of the next release.

So the check is always the same and always per artefact: under exactly which licence was this version of this model released? And if the answer is a licence bearing the company's own name — "Community License," "Research License" — you already know it is not a standard text and must be read in full, because it was written for their case, not yours.

Where to go next, with no middlemen

The Mistral 7B paper is short, public and free, and its results tables let you check exactly which tests it wins and which it does not, rather than trusting the summary. For the licence contrast, the Open Source Definition is ten points and fits on one screen.

The capability you take from this is checking the licence of each version rather than the reputation of whoever signs it — and copying quantifiers instead of summarising them, because flattening them falsifies in both directions.