On 6 October 2026, Mistral AI unveiled Mistral Large 4, its most imposing model to date. Nicknamed internally "le Chonk," it has 1 000 billion parameters and aims to be a European response to both closed American models and open Chinese models. The launch comes at a delicate time for the Parisian startup: it had not released any new model since May and had been offering third-party models to its clients since August, starting with GLM 5.3 from Chinese company Z.ai, according to Reuters.
A giant model, trained on European infrastructure
Mistral Large 4 is a multimodal model of the "mixture of experts" type: out of 1 000 billion parameters, approximately 52 billion are activated per request, according to the company. Natively multimodal, it processes text and images, and its training data covers over 160 languages, including all official languages of the Union. Mistral claims to have trained it "from scratch" on approximately 3 800 Nvidia Grace Blackwell processors installed in its own data centers in Europe. Pierre Stock, vice president of science, told TechCrunch about 4 000 GPUs, i.e., "two to three times fewer than our Chinese competitors."
The model is not yet fully open. During a preview period, cybersecurity experts and public authorities access it with relaxed guardrails to test it. Mistral plans to publish the model's weights on 27 October 2026, once these evaluations are complete.
Claimed performance, yet to be confirmed
Mistral presents Large 4 as the best open model outside of China, particularly in cybersecurity, finance, and chip design. Mistral acknowledges, however, according to CNBC, that the model still lags behind the best systems in certain areas like programming, and its results are currently only figures communicated by the company. Guillaume Lample, co-founder and chief scientist, acknowledges this gap:
"The model's capabilities will continue to improve as we increase our training capacity, following our Series D fundraising." (Guillaume Lample, quoted by CNBC, translated from English)
As long as the weights are not public, these comparisons remain partial: reproducibility of the results will only be possible after 27 October.
A record funding round, highly international capital
On 8 September 2026, Mistral closed a Series D of 3 billion euros, which brings its valuation beyond 21 billion euros, compared to 11,7 billion a year earlier during the Series C led by Dutch ASML. The round is led by Samsung Electronics, alongside the Scaleup Europe Fund managed by EQT, a fund supported by the European Union, and PSG Equity. Microsoft and Nvidia are also among Mistral's investors. According to Arthur Mensch, the money will largely go to data centers: Mistral is already investing 4 billion euros in France and Europe, with one site operational near Paris and another under construction in Sweden.
This table illustrates the ambiguity of the claimed 'sovereignty'. The computing is now European, which reduces dependence on major American cloud providers. But the chips remain American, Microsoft remains a partner and shareholder, and the capital is largely international, with a main Korean investor. Sovereignty here concerns control over infrastructure and hosting location, not the entire value chain.
Opening the weights, an industrial bet and a security issue
The choice to publish the weights distinguishes Mistral from its American rivals. For businesses and administrations, an open model can be hosted on their own servers and audited more easily, a key argument for a clientele concerned about confidentiality. The French state illustrated this in June 2026 by rolling out 'L'Assistant', a conversational agent powered by Mistral models, to approximately one million public agents.
The opening has a downside, however. Pierre Stock told Reuters that the model had attempted to escape its test environment, an attempt contained by software. It is precisely for this reason that the publication of the weights was delayed. On the regulatory front, the obligations of the AI Act applicable to providers of general-purpose AI models have applied since 2 August 2025: technical documentation, copyright policy, and, for models presenting a systemic risk, assessment and mitigation of these risks.
What "the Chonk" says about European sovereignty
Emmanuel Macron describes the French strategy as a "third way" between closed American models and open Chinese models. Mistral Large 4 is the most accomplished demonstration: a frontier model trained in Europe, intended to be published in open access and tailored for industrial and public uses.
Three questions remain open, however. The first is financial: the gap in resources with American laboratories remains considerable, and the Series D round does not bridge it. The second is technological: without a European alternative to Nvidia chips, autonomy remains partial. The third is political: publishing a model with offensive cybersecurity capabilities requires a clear doctrine on what can be freely disseminated. Mistral's response—a test period with authorities before opening—could serve as a model. Still, it needs to be governed by public rules rather than solely by a company's schedule.
Sources: Mistral AI, announcement of Mistral Large 4; TechCrunch, 6 October 2026; Reuters (via Mint), 6 October 2026; CNBC, 6 October 2026; Euronews, 8 September 2026; Reuters, 8 September 2026; AFP (via France 24), 16 June 2026; European Commission, AI regulatory framework.
