Mistral’s new 1T model aims to leapfrog closed and open rivals

French lab Mistral AI on Tuesday released Mistral Large 4, a large multimodal model nicknamed Le Chonk for its one trillion parameters. It's positioned as an alternative to both American closed models and the open models that increasingly come out of China — the 'third way' that French President Emmanuel Macron has talked about.

The notable choice is the rollout. For now the model is reachable only through a guardrailed public endpoint. Mistral says it plans to release the weights in about three weeks, after safety testing, and to work with trusted partners and governments in the meantime so the weights help defenders more than attackers.

Mistral also says the model was trained entirely on its own compute, using 4,000 NVIDIA GPUs — which its VP of science, Pierre Stock, says is two to three times fewer than Chinese competitors use, and far fewer than closed labs.

We like the shape of this release. Open weights with a safety window up front is a reasonable middle path in a year when security worries about powerful models have been rising, especially among the enterprises and institutions that make up Mistral's customer base. Stock's argument cuts both ways and we think it's right: open weights raise misuse risk, but they're also far easier to audit than a model you can only reach through an API.

The efficiency claim is the part to hold loosely. Training a trillion-parameter model on a few thousand GPUs would be a meaningful signal that the frontier isn't purely a function of who owns the most chips. But benchmark results are still pending, and leapfrogging rivals is a goal, not a result. Until independent numbers land, the fair read is that Mistral has built something big and targeted, not that it has beaten anyone.

The targeting is smart. Mistral says the model is tuned for cybersecurity, finance and chip design — and chip design isn't random. It's core to ASML, which led Mistral's Series C, and Samsung, which led its Series D last month at a €21 billion valuation. Building for the industries your backers live in is how a smaller lab finds places where it can beat bigger, general-purpose models.

The risk we're watching is the one Mistral itself names. A trillion-parameter model tuned for cybersecurity, released openly, is useful to defenders and to attackers alike. Three weeks of testing and a partner program is a start, not a guarantee. If this pattern spreads — big, capable, security-tuned open weights on a short clock — the industry will need shared norms on what that pre-release window has to prove.

Mistral recently began hosting Chinese models on its platform and has worked to show that wasn't a pivot to being a mere inference provider. Large 4 is its argument that it remains a frontier lab in its own right.