Mistral Large 4 Goes Open-Weight: Inside the 1 Trillion Parameter Model

Mistral has just made its biggest model public. The French AI lab released the open weights for Mistral Large 4 — nicknamed “Le Chonk” — a 1-trillion-parameter multimodal model that first appeared as a public preview API on October 6, 2026. For developers and researchers who’d rather run frontier-class AI on their own hardware than rent it through an API, this is the release of the month.

It’s a big deal for a simple reason: models at this scale have almost always stayed locked behind corporate APIs. An open-weight trillion-parameter model means anyone with enough compute can download it, fine-tune it, audit it, and deploy it where they want — including places a vendor’s cloud never reaches.

What Mistral Large 4 actually is

According to Mistral’s announcement, Large 4 is a mixture-of-experts (MoE) model with roughly one trillion total parameters and 52 billion active parameters per token. It’s natively multimodal, so it reasons across text and images rather than bolting vision on as an afterthought. Mistral describes it as an open-weight hybrid instruct-and-reasoning MoE, built to unify instruction-following, reasoning, coding, and agentic behavior in one model.

That combination matters. Instruction models answer questions; reasoning models think through problems; agentic models use tools and produce finished work. Large 4 is meant to do all three, which is why Mistral positioned it for everything from software engineering to financial and legal analysis.

From preview to open weights: the timeline

Mistral unveiled Large 4 on October 6, 2026, launching a public preview API through Mistral Studio while promising the weights would follow by the end of the month. The company spent the interval red-teaming the model before the open release — a sensible precaution at this scale. The open weights landed in late October, right on schedule, completing the rollout the preview had started.

If you followed our earlier coverage of the October AI price war, you’ll remember Large 4 debuted with API pricing of just $0.70 per million input tokens and $2.10 per million output tokens — aggressively cheap for a model in this class. Now the weights themselves are available to download and run anywhere.

What it’s good at

Mistral’s report card for Large 4 focuses on practical, professional work rather than trivia benchmarks:

• Coding: strong results on software engineering evaluations like DeepSWE 1.1 and Terminal-Bench 4.0, plus a blind human evaluation of coding quality run with Surge AI.
• Agentic workflows: the model gathers information, uses tools, and produces finished deliverables like spreadsheets and reports.
• Cybersecurity: among the top scores on the Artificial Analysis Cyber Index — notable because provider-level content refusals in closed models can block legitimate defensive security work.
• Multimodal understanding: a clear step forward in image reasoning for Mistral, with strong visual grounding across charts, documents, and complex images.
• Knowledge work: third-party evaluations on finance and legal tasks put it ahead of models like GPT-6-Astra on representative professional tasks.

How it was built

Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral’s own datacenters in Europe — part of the company’s pitch that this is, in its words, “forged in Europe” for AI sovereignty. The reinforcement learning setup runs at serious scale: around 3,000 GPUs generating tens of thousands of rollouts in parallel, producing roughly 33 billion tokens per day during training.

The whole program sits on the roadmap funded by Mistral’s €3 billion Series round, and the company says more compute is already on the way — so Large 4 is a milestone, not a finish line.

Why the open release matters

Open weights change the power dynamic. Enterprises can deploy Large 4 in their own regions, including the European deployment Mistral offers, keeping sensitive data in-house. Security teams can run it without worrying that a vendor’s refusal filters will block legitimate defensive research. And researchers can finally study, reproduce, and build on a trillion-parameter MoE — the kind of artifact that usually stays behind closed doors.

It’s also a statement of intent. While most frontier labs keep their largest models proprietary, Mistral keeps betting that open-weight releases build the ecosystem that wins in the long run. With a trillion parameters now downloadable, that bet just got a lot more interesting. The official docs have the full model card for anyone who wants the details.

For a look at the hardware this kind of model runs on, see our piece on AMD’s trillion-dollar chip ambitions.