Mistral Large 4 launches with 1T parameters, open weights promised for end of October

Mistral launches a trillion-parameter multimodal model

French AI lab Mistral AI has released Mistral Large 4, a natively multimodal model with one trillion total parameters and 49 billion active parameters per inference, according to the company's announcement on X. The model, nicknamed "Le Chonk" in a nod to its size, is available to all customers through Mistral's API as of Tuesday, but its weights will not be published until the end of October.

According to TechCrunch, the launch is an attempt to leapfrog both American and Chinese rivals, in line with what French President Emmanuel Macron has described as "a third way in AI." Mistral is positioning the model as an alternative to closed US frontier models that can be shut off at the provider's discretion, and to open-weight models that increasingly come out of China.

What Mistral claims

In its announcement, Mistral calls the model the best open-weights model from the US or Europe on aggregated benchmarks, and claims state-of-the-art results on critical workloads including cyber defense, manufacturing and finance. The company also says it surpasses closed frontier models on visual grounding, the task of identifying and localizing objects within images.

TechCrunch reports that independent benchmark results were still pending at publication time, and that Mistral framed its ambition as becoming best in class among open-weight models, particularly outside China.

Not open weights yet

Despite the open-weights framing, the model cannot currently be downloaded. TechCrunch says it is reachable only through a public guardrail endpoint, with weights to follow in about three weeks, once safety testing is complete.

"In the meantime, we'll work with trusted partners and governments to make sure that the open-source weights can be used to defend, but not to [perform] malicious attacks," Mistral VP Science Pierre Stock told TechCrunch. Stock argued that open-weight models are easier to audit, which he framed as an advantage for Mistral's core audience of enterprises and institutions at a moment when security concerns are mounting. Mistral's own announcement adds that it is working with cybersecurity partners privately in the interim.

Training economics and target domains

Stock told TechCrunch the model was trained entirely on Mistral's own compute, using about 4,000 NVIDIA GPUs. He contrasted that figure with what he described as two to three times more compute at Chinese competitors, and significantly more at closed-source labs.

According to Stock, the model's optimized use cases include cybersecurity and finance, but also chip design. TechCrunch points out that chip design is core to two of Mistral's main backers: ASML, which led its Series C, and Samsung, which led a Series D round last month at a €21 billion valuation, roughly $24.39 billion.

A sovereignty pitch

Mistral's announcement stresses that the model was "forged in Europe end-to-end" and is deployable from Europe through the company's own Mistral Cloud infrastructure. TechCrunch adds context to the launch: Mistral had recently begun hosting Chinese models on its platform, prompting suggestions that it was drifting toward becoming a mere inference provider. With Mistral Large 4 in its portfolio, the company argues it should still be regarded as a frontier lab.

Why it matters

Mistral is betting that a European lab can stay at the frontier without sealing its weights behind an API or matching the compute budgets of the largest US and Chinese players. If the benchmark claims hold up under third-party testing, a frontier-class open-weights model trained and hosted in Europe would give enterprises a sovereignty-friendly alternative to US cloud APIs, and it would complicate the assumption that openness costs performance. The staged release is also a live experiment in how "open" gets defined: weights are promised, but only after governments and trusted partners have had a say in how they can be used. Until independent evaluations arrive, the model's frontier credentials rest on Mistral's own claims.