YFarmX logoYFarmX

Mistral AI

Mistral Large 4

Downloadable weights and hosted access for a trillion-parameter model

Weights released 29 October 20263 min readLarge Language ModelsLast updated:

Editorial illustration of a Mistral-branded computing module beside a photograph and code sheet, headed Mistral Large 4.

Key facts

1.05T
Total parameters
1M tokens
Context
API + weights
Access today
52B
Active parameters

Mistral Large 4 is an AI model that reads words and pictures, writes answers and code, and works through software tools. Mistral opened its public preview on 6 October 2026.

Listen to the 6 October preview guide

Podcast: Mistral's trillion-parameter preview

Mistral Large 4 is an AI model that reads text and images, writes answers and code, and takes actions through software tools. Mistral opened its public API preview on 6 October 2026, with downloadable weights released on 29 October.

Mistral released the model weights on 29 October. These are the learned numerical settings needed to run the model on suitable infrastructure. The model repository lists 1.05 trillion total parameters and 52 billion active parameters.

What can Mistral Large 4 do?

Mistral Large 4 combines image understanding, coding and tool use in one model. Its model documentation lists a one-million-token context window, function calling, structured outputs, document questions and conversational agents. Tokens are the pieces of text and encoded content the model processes.

A tool-using assistant can inspect a document, ask an application to retrieve information and use the result in its answer. The application decides which tools it exposes and which actions require a person’s approval.

Mistral’s launch demonstrations include analysing financial filings, examining technical drawings and investigating software vulnerabilities. These examples describe the tasks the company is targeting. A useful evaluation uses the documents, codebase and tool permissions the intended deployment will actually handle.

How does its trillion-parameter network work?

Mistral’s model page lists 1.05 trillion total parameters and a mixture-of-experts architecture. Parameters are the numerical settings learned during training. A learned router selects expert networks for each token, as Nvidia’s explanation of the architecture describes.

That selection lets a model carry a large collection of learned knowledge while using selected parts of its network at each step. The expert outputs are combined before processing continues through later layers.

Hosting still requires memory for the weights and efficient communication between processors. Total model size, the work activated at each step and the length of the input all affect deployment requirements. A hosted API lets a team evaluate results before planning its own serving infrastructure.

Schematic mixture-of-experts layer: a token passes through attention and a learned router, selected expert networks process it, and their outputs combine.
A learned router selects expert subnetworks for each token. This is a schematic explanation of a mixture-of-experts layer. Sources: Mistral and Nvidia.

What does hosted access cost?

Mistral’s 29 October release lists the following API prices, checked on 9 November 2026. Each figure is in US dollars per million tokens.

Token category Published API price
Input $0.70
Cached input $0.07
Output $2.10

Input is the content sent to the model; output is the content it generates. Cached input is previously processed content reused under the service’s caching rules. The total charge depends on the quantities billed in each category.

The hosted model uses the identifier mistral-large-4. Reasoning can be set to low, medium or high, with high as the default. Downloadable weights use the Mistral Ecosystem Open-Weight licence, which has separate commercial terms for businesses offering models as a service.

How should you read its benchmark scores?

Mistral’s launch post reports the following selected results. They describe separate evaluations with different tasks and scoring rules.

Evaluation Reported score Task
DeepSWE v1.1 61.7% Software engineering
Terminal-Bench 4 28.3% Terminal workflows
Cybench 93% Security challenges

These are Mistral’s published 6 October preview figures. The company attributes its DeepSWE and terminal results to Artificial Analysis; Cybench covers a set of security exercises.

For a deployment decision, run the same representative tasks through the models being compared. Keep tool access, prompts and output budgets consistent, and record completed tasks, errors, elapsed time and token charges. That produces a comparison tied to the work the model will do.