Mistral Large 4 explained: Europe's 1-trillion-parameter open-weight model

On 6 October 2026, the French AI company Mistral opened Mistral Large 4 in preview through its API, and promised the model's open weights for 27 October. It is a roughly 1-trillion-parameter model, trained from scratch in Mistral's European data centres, and Mistral calls it the most powerful open-weight AI system outside China. For European companies that need to keep data in the EU, or to run a model on their own servers, it is the most important release of the autumn.

What Mistral built

Large 4 reads text and images and writes text. Like most frontier open models, it uses a design where only a small part of the network, about 49 billion parameters, works on each token, which keeps it fast and affordable to serve. Mistral trained it from scratch in about two months on roughly 4,000 Nvidia Grace Blackwell GPUs in its own European data centres, drawing about 10 megawatts. Reinforcement learning is still running, and Mistral says the gains have not slowed, so the final model may be stronger than today's preview.

Benchmarks (preliminary)

All figures below are Mistral's own and marked preliminary; it expects them to change before the weights ship. They focus on practical, agent-style tasks rather than exam-style questions.

It also leads Mistral's visual grounding tests on satellite images (73% versus 68% for GPT-6 Astra). On the independent Artificial Analysis Intelligence Index, the preview scores 38, well behind the top closed models and the best Chinese open models. In short: very competitive on the agentic and business tasks Mistral targeted, not yet a frontier all-rounder.

Pricing

That is a fraction of the top closed models, and output is priced especially low, which matters for long answers and agent loops. For a side-by-side view of prices and quality, see the best AI models in October 2026.

Why it matters for European data teams

Mistral says developers, cybersecurity firms and government agencies will get a version with fewer restrictions and extra cyber features. It already works with more than 125 large businesses, including Airbus, ASML and HSBC. If you are tracking the EU AI Act's obligations for general-purpose models, an open European model with published documentation is likely to be easier to assess than one with little detail.

More on open models: Meta's Muse Spark Open, the other big open-weight launch this month. For the rules that apply to AI models in Europe, see the EU AI Act timeline.

Sources: The Next Web, "Mistral releases Large 4" (6 October 2026); ODSC's AI news roundup for 5 to 11 October 2026 (API pricing); the Artificial Analysis leaderboard (mid-October 2026). All benchmark figures are Mistral-reported and preliminary.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 17 October 2026. About me