Mistral Large 4 open weights: what Europe's 1-trillion-parameter model means for SMEs

Mistral released the open weights of Large 4. What the European model offers, its real limits and how SMEs can use it with more control over data.

On 29 October 2026, Paris-based Mistral AI published the weights of Mistral Large 4, a model the company nicknames "Le Chonk". It had been available as a paid API preview since 6 October; now anyone can download it from Hugging Face, run it on their own infrastructure and adapt it. For European companies that have been asking how to use powerful AI without depending entirely on a handful of US providers, this is the most relevant AI news of the week.

What was released

According to Mistral and the official model card, Large 4 is a natively multimodal mixture-of-experts model with about 1 trillion parameters, of which roughly 49 billion are active for each token (52 billion including embeddings and output layers). It combines an instruct mode and a reasoning mode with three effort levels, supports a context window of up to 1 million tokens and is described as fluent in more than 160 languages. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters.

The weights are published under the "Mistral Ecosystem Open-Weight" licence, which Mistral describes as a modified MIT licence: use, modification, fine-tuning and commercial use are allowed, but companies offering the model itself as a service (Model-as-a-Service) need a separate commercial licence. For those who prefer not to host it, the model is available through the Mistral API at $0.70 per million input tokens and $2.10 per million output tokens, as well as via OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry.

How good is it?

Mistral positions Large 4 as the strongest open-weight model for agentic coding and cybersecurity. Among the figures it reports: 48% on the Artificial Analysis Coding Agent Index, "effectively tied" with China's GLM-5.3 while using on average 47% fewer tokens per task, and 41% on Terminal-Bench 4.0, up 13 points from the preview. These are vendor-reported numbers, and independent coverage is clear that the leading proprietary models from OpenAI and Anthropic remain ahead overall. Fortune also reports that Artificial Analysis found the preview more expensive per task than open models of similar intelligence. In short: a serious frontier-class open model, not the best model in the world.

Why it matters for European companies

The main argument is not raw benchmark scores but control. In Fortune's interview, Mistral's VP of science Pierre Stock said that "business continuity" is the top request from customers in regulated industries, who worry that access to closed models "can be cut off". With open weights, a company or its IT provider can keep a copy of the model and run it in an EU datacenter, a private cloud or on premises. That has practical consequences:

  • Data residency and confidentiality: sensitive documents, contracts or customer data can be processed without leaving infrastructure you control, which simplifies GDPR discussions with clients and auditors.
  • Vendor independence: prices, terms or availability of an API can change; a model you host keeps working.
  • Customisation: the licence allows fine-tuning on your own terminology, products and procedures.
  • Long documents: a 1-million-token context makes it realistic to analyse large manuals, tenders or code bases in a single pass.

A reality check for SMEs

"Open" does not mean "runs on a laptop". The model card's deployment examples assume a node of eight high-end GPUs with vLLM or SGLang; a quantised NVFP4 checkpoint reduces the memory footprint but targets NVIDIA Blackwell hardware. For most small and medium-sized businesses, the realistic paths are:

  1. Use it via an EU-hosted API (Mistral operates a European deployment end to end) and keep the option of self-hosting later.
  2. Ask your cloud or IT partner to host it on dedicated EU infrastructure if confidentiality requirements justify the cost.
  3. Choose smaller open models for everyday tasks such as drafting, classification or internal search, and reserve a frontier model for the work that really needs it.

Whatever you choose, the basics do not change: decide which data may go into which tool, test the model on your own real cases before rolling it out, keep a human in the loop for decisions that affect customers or employees, and make sure your people know how to use AI responsibly. The EU AI Act also asks every organisation to understand its role (provider or deployer) for each AI system it uses.

Our take

Mistral Large 4 does not end the dominance of US and Chinese labs, but it gives European companies a credible, frontier-class option they can own and audit. The smartest move for an SME is not to switch everything overnight, but to design its AI processes so that the model behind them can be replaced: clear prompts and workflows, documented data rules and a short list of tested alternatives.

This is exactly the kind of choice we help teams make in the AI Evolution Program: mapping where AI really adds value, selecting tools that fit your data and compliance needs, and training people to use them well. If you want to evaluate open models such as Mistral Large 4 for your company, get in touch.

Want to bring these ideas into your company?

Let's talk →