Mistral Large 4 Arrives: 1 Trillion Parameters, 49B Active

French AI company Mistral has made Large 4 generally available, describing it as the best open-weight model on the market. The company has nicknamed the model "Le Chonk," a nod to its sheer size: 1 trillion parameters in total.

That size does not mean the whole model runs on every request. According to Mistral, Large 4 activates only about 49 billion of its parameters at a time, depending on the task. The model is also multimodal, so it can process images as well as text.

1 trillion parameters, 49 billion active

The most notable feature of Large 4 is the gap between its total parameter count and the number of parameters engaged during processing. By the company's own figures, the model is 1 trillion parameters in size, while the active portion stands at 49 billion.

Running only a fraction of a model's parameters on each pass is a common approach aimed at delivering the capacity of a very large model with a more manageable computational load. Parameters are the numerical values a model learns during training and which determine its outputs. They are often used as shorthand for a model's scale, though the count alone does not guarantee quality.

Bold claims on critical workloads and visual grounding

Mistral says Large 4 performs very well on critical workloads such as cyber defense, manufacturing and finance. The company also claims that, on aggregate benchmarks, it is the best open-weight model from the US or Europe.

Another claim concerns visual grounding. According to Mistral, Large 4 can outperform rivals' closed frontier models in this area. These assessments rest on the company's own statements. Multimodal capability means the model is not limited to text, but how it performs in practice will become clearer as users put it through its paces.

Coding improvements and a European angle

The final open-weight release published today includes improvements that the company highlights particularly for coding. Mistral also stresses that Large 4 was developed entirely in Europe.

Being open-weight means the model's weights are available to developers and organizations, who can run it on their own infrastructure and adapt it to their needs. The announcement can be read as an effort to put a European option forward in the open-weight space. Whether it succeeds will depend on how well the company's claims hold up under independent evaluation.