Mistral Large 4: Inside the Le Chonk Preview
Mistral Large 4 has entered public preview under the nickname Le Chonk. Mistral announced the multimodal model on October 6, 2026, with open weights planned for the end of October. Developers should distinguish the available API preview from the future weight release. Mistral’s announcement describes the launch and its European infrastructure.
Event date: October 6, 2026 · Sources checked: October 7, 2026
What the Mistral Large 4 announcement confirms
The company describes a sparse model with roughly one trillion parameters. Its announcement cites 49 billion active parameters per token and training on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. These are company-reported specifications, rather than measurements independently reproduced by xpu live.
There is a specification difference worth preserving: the current model documentation lists 1.05 trillion total parameters and 52 billion active parameters, alongside a one-million-token context window. Readers should use the current technical documentation when planning deployment instead of treating rounded announcement figures as final implementation requirements.
Preview access and open weights are different milestones
Mistral’s release notes identify the model as a public preview. They also describe a temporary launch discount, which means a quoted promotional rate should not be mistaken for a permanent tariff. The planned weight release remains a future event at the time of this article.
The official changelog is the appropriate place to recheck availability. A preview can evolve as the provider changes serving settings, model behavior or access conditions.
Mistral Large 4 — xpu live analysis: evaluate the workload, not the size
A large parameter count alone cannot establish whether a model is the best choice for an application. Our editorial view is that teams should test representative text and image tasks, record failures and compare the complete request cost. A long context window also does not guarantee reliable recall throughout a long document.
For a potential self-hosted deployment, separate the future license and weight availability from hardware feasibility. Memory requirements, serving software and actual throughput will matter. Keep the preview model identifier in your test records so a later release can be evaluated against the same workload. The useful next milestone is inspectable weights and reproducible testing, rather than another headline number.
Sources and further reading
Related on xpu live: Model releases evaluation checklist.