Mistral ships Large 4: a trillion-parameter open-weight model trained entirely in the EU
Mistral unveiled Large 4 on October 6, a 1.05 trillion parameter open-weight mixture-of-experts model trained on 3,800 Nvidia GB200 units inside its own EU data centers. Weights are promised later in October.
The decision it puts on your desk
Price a pilot on Large 4's weights the week they land, later in October. If EU data residency is a requirement, benchmark it against your current provider before the EU procurement window closes. If you sell compute inside the EU, re-price your capacity offers for trillion-scale training and inference demand, because Mistral's own centers have now proven the market exists.
Mistral's fall flagship landed at trillion-parameter scale on Tuesday, and geography is part of the release's argument.
Large 4 is a 1.05 trillion parameter multimodal Mixture-of-Experts (MoE, a design where only the relevant parts of the model wake up per answer) model, with 49 billion active parameters and a one million token context window.
The training case sits in the release's own numbers: 3,800 Nvidia Grace Blackwell GPUs, running inside data centers Mistral owns in the European Union.
Weights arrive later in October, the company promised.
What a trillion-parameter open weight is for
Ecosystem check: Mistral is the European entry in the trillion club, and the open-weight tier now holds this scale.
Chinese labs released trillion-parameter weights this month too, so the market for self-hostable large models carries at least two credible options at the top tier.
Under the hood, the active-parameter design keeps inference affordable at scale. Only 49 billion of 1.05 trillion parameters run per token, which cuts the per-answer compute cost that dense trillion-parameter models carry.
Precedent pricing says the release also supports the September round. The $24 billion post-money valuation closed weeks ago, and a trillion-parameter EU-trained model is the asset pitch for it.
The unresolved edge is benchmark reality. Launch claims stay vendor-conditioned, and downloaded weights get rerun by outsiders after October's weight drop.
I get the sense the buyer type is named: European governments and enterprises that want frontier-class scale inside EU infrastructure, without routing data or spend through American providers.
Source
← Previous
Meta and Sierra published the Personal Agent Protocol. Payments rails and retailers just joined an agentic standard
Next →
Anthropic launches Claude Haiku 5.5 at $0.10 per million tokens, and cuts Sonnet's cache reads
The Morning Recap
We sift the day. You make the calls.
The decision on every story, in your inbox before your first call.