Mistral Large 4: Europe's 1T-Parameter 'Third Way' in AI

Every few weeks a new AI model arrives with a big number attached and a promise to change everything. This one deserves a slightly closer look, and not only because of the number. On Tuesday, French lab Mistral AI released Mistral Large 4 (ML4), a large multimodal model with roughly one trillion parameters. According to TechCrunch's report, Mistral wants it to leapfrog both American and Chinese rivals.

What makes the story interesting is less the size and more the positioning. Mistral is trying to sit in a gap that has opened up in the AI market, between closed models that can be switched off and open models that mostly come from China. Let's unpack what that means, what Mistral has actually said, and what is still missing.

What Mistral actually released

ML4 is a multimodal model, meaning it is built to work with more than plain text. Mistral has nicknamed it "Le Chonk", a nod to its one trillion parameters. Parameters are, roughly speaking, the adjustable numbers inside a model that get tuned during training. More of them usually means more capacity to learn, although it also means more cost to train and run, and bigger does not automatically mean better.

There is a catch worth flagging early. ML4 is not an open-weight model yet. For now it can only be reached through a public endpoint with guardrails, which is a managed access point where Mistral controls what goes in and out. The company says it plans to publish the weights in about three weeks, once safety testing is complete.

That sequencing tells us something. Open-weight releases are essentially irreversible: once the weights are out, anyone can download, modify and run the model, and nobody can recall it. Testing first and releasing second is the cautious order of operations, and in a year when security worries about powerful models are loud, it is a sensible one.

The 'third way' between closed and open

TechCrunch frames the release against a growing divide. On one side are closed models, the kind you can only use through someone else's servers, which the article notes can be unplugged. That link points to a June story about dozens of cybersecurity experts urging the White House to lift export-control restrictions on Anthropic's most powerful models, arguing that the restrictions would hamper defenders. Whatever you think of that dispute, it illustrates the risk: if your business depends on a model you do not control, a policy decision in another country can change what you are allowed to use.

On the other side are open models, which TechCrunch says are often made in China. They give users control, but many European enterprises and institutions may hesitate over where a model comes from, even when it is technically inspectable.

Mistral's pitch is that ML4 offers a third option: a frontier-scale model from a European company, with open weights coming. French president Emmanuel Macron has used almost exactly this language, describing "a third way in AI", and TechCrunch notes that ML4 follows that framing. It is a politically convenient slogan, but it also describes a real commercial need that customers have.

Security: the double-edged sword of open weights

One of the most candid parts of the story is about security. Mistral's VP of Science, Pierre Stock, told TechCrunch that while waiting for the weights release, the company will work with trusted partners and governments so the model can be used to defend rather than to attack.

This is the central tension of open models in cybersecurity. A strong model can help a defender find vulnerabilities in code before criminals do. It can help an attacker do the same thing. Once weights are public, you cannot decide who gets which version of that capability.

Stock also made the counterargument: an open-weight model is easier to audit. If you can inspect and test a model yourself, you do not have to take a vendor's word for what it does. For governments and regulated companies, that auditability can matter as much as raw performance. If you want more background on how AI is reshaping defence and data protection, our piece on AI and cybersecurity covers the basics.

My own view: neither side of that argument wins outright. Auditability is real, but auditing a trillion-parameter model is hard, and "easier than a closed model" is a low bar. The honest summary is that open weights shift responsibility toward the people deploying the model, which is exactly why a testing window before release is welcome.

Doing more with less compute

Another notable claim concerns how the model was trained. According to Stock, ML4 was trained entirely on Mistral's own compute, using only 4,000 Nvidia GPUs. He described that as two to three times fewer than its Chinese competitors use, and significantly fewer than closed-source competitors.

If that holds up, it matters for two reasons. First, it suggests efficiency: a smaller training fleet is cheaper and requires less energy. Second, training on your own hardware is a sovereignty statement, since it means no dependence on a cloud provider for the most sensitive step of building a model.

I would treat the comparison with some caution, though. It comes from a Mistral executive, and the article does not give figures for rivals that we can check. It is a claim to watch, not a verified fact. Compute counts are also only part of the story: training time, data quality and techniques all affect what a model can do.

Benchmarks are still pending

Here is the part that keeps the excitement in check: benchmark results are still pending. Mistral hopes ML4 will be the best in class among open-weight models, especially outside China, though Stock suggested the ambition goes beyond that. Because of its focused training, the company thinks ML4 could outperform closed models in specific areas that matter to its customers, where multimodal abilities add value.

That wording is careful, and I think it is telling. Mistral is not claiming to beat every closed model at everything. It is claiming to be best among open-weight models and to win in particular niches. That is a more defensible and more honest position than a blanket "state of the art" claim, but it also means we cannot judge the model until independent numbers appear.

Until then, "leapfrog" is an aspiration. It is Mistral's goal, as the headline says, not a result.

Chips, finance and security: the target use cases

According to Stock, the use cases ML4 has been optimised for include cybersecurity and finance, as well as chip design. That last one is not random. Chip design is core to two of Mistral's main backers: Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month at a valuation of €21 billion (about $24.39 billion).

Finance is another natural fit for a company selling to institutions that care about control and compliance. If you are curious how AI is already changing banking, we explored it in our article on AI in financial services.

Multimodality matters in these fields because the work is often not just text. Think of diagrams, layouts, charts and scanned documents. A model that can read and reason across those formats can be more useful in a specialised workflow than a text-only system, even if it is not the top performer on general trivia.

The money and the strategy behind it

To understand ML4, it helps to look at the September funding round. TechCrunch reported that Mistral raised €3 billion at a post-money valuation of more than €21 billion, in a Series D led by Samsung Electronics, with the EQT-managed Scaleup Europe Fund and existing investor PSG Equity as co-leads. Mistral called it the largest equity round ever completed by a European technology company.

The company said it would use the money to scale compute, build infrastructure, accelerate commercial growth and expand internationally. It also wants 1 GW of compute capacity in Europe by 2030. In August it launched tools that let customers choose in which regions their AI queries are processed, and it has started hosting third-party open-weight models, including Chinese ones.

That last move drew criticism. Some read it as a sign that Mistral was becoming a mere inference provider, a company that rents out other people's models. TechCrunch says Mistral pushed back, calling its frontier research the foundation of its infrastructure, products and sovereignty. In that light, ML4 reads as a proof point: we still build our own frontier models.

The investor list is also quite international. Samsung leads, but American names such as a16z, Nvidia and Salesforce Ventures took part, along with new backers Advent and BlackRock. Microsoft's partnership with Mistral was significantly expanded in July. So the "third way" is not about cutting ties with the US. It is about not depending on any single power.

What this means for the rest of us

If you are not buying AI for a bank or a chipmaker, why should you care? Because the question behind this launch is one that touches everyone: who controls the AI that more and more of our services run on? If a handful of companies in one country hold all the powerful models, then a policy change, a price hike or a shutdown can ripple into the apps and public services we use daily.

A more varied ecosystem, with credible European options and auditable open weights, could spread that risk. But variety is only valuable if the models are actually good, safely released and honestly evaluated. We do not know yet whether ML4 clears that bar. The three-week wait for the weights and the pending benchmarks will tell us much more than any launch-day claim.

I would keep an eye on three things: independent benchmark results, the conditions under which the weights are released, and whether Mistral's efficiency claim survives outside scrutiny. And if you use AI tools at work or at home, it is worth asking where your tools come from and what happens if they disappear tomorrow. How much would you want to know about the model behind the products you rely on, and would a European or open alternative change your mind?