Mistral Large 4: What to Know About the AI Model

What is Mistral Large 4?

Mistral Large 4 is a general-purpose artificial intelligence model from Mistral AI. The company describes it as an open-weight, multimodal model. “Multimodal” means it can work with more than one kind of input: in this case, text and images. Artificial Analysis lists its output as text. ¹²

A model is the underlying system that an AI service uses to interpret a prompt and produce a response. The supplied sources describe the model and some evaluation results, but do not explain which products or interfaces make it available to individual users.

What does its design mean?

Mistral describes Large 4 as using a granular Mixture-of-Experts architecture. In plain language, this is a design in which a model can route work through selected parts, or “experts,” rather than using every part for every request. Mistral reports 1.05 trillion total parameters and 49 billion active parameters, along with a 1.6-billion-parameter vision encoder for image-related processing. ¹

Those figures describe the model’s design, not a guarantee that it will answer well or quickly in every situation. The source does not provide enough detail to explain precisely how the routing works or how it affects a particular user’s hardware, cost or results.

Can Mistral Large 4 understand images and long prompts?

The available model information says Large 4 accepts text and images and produces text. ² That makes image questions one described use, but the sources do not give examples or specify which image tasks it handles reliably.

Artificial Analysis lists a context window of 524,000 tokens for Mistral Large 4 Preview. ² A context window is the amount of text and other supported input a model can consider at once. A large window may be useful for long documents or extended conversations, but it does not mean the model will remember information forever or use every detail correctly.

How does it perform in tests?

Artificial Analysis gave the Preview a score of 38 on its Intelligence Index and said that was above the median of 26 for its comparison group. It also reported an average speed of 116 tokens per second and described the model as notably fast. ²

These are results from one evaluator’s tests, not a universal rating. The index combines several evaluations, and results can differ by task, prompt and comparison set. The evaluator also noted that the model generated substantially more tokens than the median during its index testing, which it described as very verbose. ² In practical terms, a lengthy answer may need editing, and the reported speed alone does not establish how quickly every user will receive a response.

What does Mistral Large 4 cost, and what is still unclear?

For the Preview, Artificial Analysis lists prices of $1.36 per million input tokens and $4.18 per million output tokens. It also estimates an average cost of $1.13 per task in its Intelligence Index evaluation. ² These are evaluator-reported figures for that version and setting; they do not establish what every service will charge or what an individual’s usage will cost.

The sources do not establish how the model performs across everyday tasks, whether the Preview figures match a final or later version, or what access options are available as of October 7, 2026. Mistral calls the model open-weight, but the supplied material does not spell out the specific terms for access or use. Readers should check Mistral’s current model documentation for the latest details. ¹