Mistral has finally shipped a frontier-class model, and it is a statement piece. Large 4, immediately nicknamed "Le Chonk" by an internet that could not resist, is a natively multimodal mixture-of-experts model with 1.05 trillion total parameters and 49 billion active parameters, wrapped around a 1-million-token context window. It was trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters — no rented American cloud, no borrowed infrastructure — and it is available right now as a public preview through Mistral's API. The open weights, the part the community actually cares about, are promised for the end of October, pending a safety review. For a company that built its reputation on efficient open models punching above their weight, this is a deliberate step into the heavyweight division.
The early benchmark picture is uneven in an interesting way. Mistral's numbers are strongest on vision, where Large 4 scores 42% on Dense 200, edging out GPT-6 Astra's 41%, and on cybersecurity, where it clears 82% on CyberGym-E2E and beats the leading Chinese open-weight models head to head. On broader general reasoning it still trails the frontier, and Mistral is not pretending otherwise. One data point worth more than the press release: a Plotly engineer running the company's internal data-analytics benchmark reported accuracy jumping from 58% to 74% at ten times lower cost than Mistral Medium 3.5 — a "generational shift," in his words. Independent replication will matter here, because the gap between a vendor benchmark table and production behavior is where AI projects go to die.
The framing around the launch is as notable as the model itself. Mistral is explicitly selling European sovereignty: with US export restrictions limiting OpenAI and Anthropic model availability abroad, European demand for a homegrown frontier alternative has sharpened considerably, and TechCrunch reports Mistral trained Large 4 on two to three times fewer GPUs than its Chinese competitors needed for comparable efforts. The enterprise pitch targets three verticals — cybersecurity, finance, and chip design — the last one notable given strategic backing from ASML and Samsung, who have obvious interest in AI-assisted silicon workflows. Wired's coverage frames the release as Europe's bid to matter in a race that has looked increasingly like a US-China duopoly. The timing carries some irony, though: the same week, Ecosia's CEO told Politico the Berlin search engine was dropping Mistral for open-weight models, saying "we are disappointed with the quality of Mistral" — a reminder that sovereignty narratives do not survive contact with bad evals.
For practitioners, the thing to watch is the end-of-October weights drop. If Large 4's weights land under a permissive license with the cyber and vision chops intact, it becomes a serious self-hostable option for security tooling, document-understanding pipelines, and multimodal agents — and Europe's first credible answer to the Chinese open-weight wave. Until then, treat the benchmark table as a claim rather than a measurement, run your own evaluations on the API preview, and pay special attention to the reasoning gap on your specific workloads. The model to evaluate is the one you can download, not the one in the launch blog post.
