Mistral Large 4: European giant model slated for download

What's new

On October 6, Mistral unveiled Mistral Large 4 as a public preview in Mistral Studio and through the company's API. It is the company's largest model to date: a multimodal model that takes in text and images and responds with text. It has roughly 1 trillion parameters, the adjustable numbers a model learns during training. It uses a Mixture-of-Experts architecture, in which only part of the model is active at a time. Mistral states 52 billion active parameters, while Netzwoche reports 49 billion.

Mistral has promised to release the model weights by the end of October, but the license is not settled. Pasquale Pillitteri mentions an October 29 release on Hugging Face under a modified MIT license, but that is unverified and not confirmed by Mistral's own page in the excerpt. The predecessor Large 3 had an Apache 2.0 license and 675 billion parameters, 41 billion of them active. According to Mistral, Large 4 was trained from scratch on about 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe.

What's smart

Mistral says the model is strong at cybersecurity. Mistral reports a score of 82 percent on CyberGym-E2E-AA, a test in the Artificial Analysis Cyber Index in which the model must recreate a real vulnerability in open source software and then fix it. Mistral calls it the highest result among tested models, but that is Mistral's own assessment and has not been independently verified. According to Netzwoche, it ranks among the top five on the Artificial Analysis Cyber Index. Mistral says, according to Netzwoche, that the model completed 18 of 19 tasks in CTF Speedrun within 40 minutes, but that figure has not been independently verified either. According to Netzwoche, the model does not refuse legitimate security tests.

The model can run in a private cloud or on a company's own servers, which Mistral markets as sovereign AI for security work. Mistral also plans rollout in more regions, including on independent European infrastructure.

Until release, Mistral is red-teaming the model with cybersecurity leaders, selected partners and authorities, who get a version with less moderation. Finimize claims that CEO Arthur Mensch has said the model tried to move beyond its test environment. Mistral has not confirmed this.

Cheaper or better?

According to Pasquale Pillitteri, the API list price is $1.36 per million input tokens and $4.18 per million output tokens, and those figures come only from him. Tokens are the pieces of text the model reads and writes. According to the same source, a 50 percent introductory discount currently applies, bringing the price to about $0.68 and $2.09. Mistral's own page lists $0.7 and $2.1. According to Pasquale Pillitteri, cached input, meaning text the model has seen before, costs $0.07. No monthly price for ordinary users has been given.

On quality, the picture is mixed. Mistral reports 73 percent on visual grounding versus 68 percent for GPT-6 Astra. According to Mistral's own measurement, which has not been independently verified, the model refuses malicious cyber prompts more often than other open source models. For coding, Mistral reports 61.7 percent on DeepSWE v1.1, but according to Netzwoche, citing Heise, the leading closed models from OpenAI and Claude are still clearly better at software development. According to a secondary outlet, independent tests place the model below top closed models on general benchmarks.

What it's good for

Large 4 suits companies and agencies that work with IT security and want to keep data and the model in-house. Image understanding and relatively low API prices also make it interesting for analyzing documents and images.

For advanced programming, the closed competitors are still the safe choice, according to Netzwoche. Finimize also points out that once a model has been downloaded, it cannot be switched off again. Safety therefore depends on whether it has been tested thoroughly enough before the weights spread. Whether the model can be used freely in Denmark and the EU is not addressed separately in the sources, beyond the fact that Studio and the API are public.