Visit Tool

Mistral Large 4 is a new multimodal AI model from Mistral AI with 1 trillion total parameters and 49 billion active parameters. A public preview is available through Mistral Studio, while Mistral plans to release the model weights by the end of October 2026.

The model is designed for coding, cybersecurity, business automation, scientific reasoning, document analysis, and visual tasks. Mistral says it trained Large 4 from scratch using 3,800 NVIDIA Grace Blackwell GPUs in its European datacentres.

What Is Mistral Large 4?

Mistral Large 4 is a natively multimodal artificial intelligence model built to handle different types of workloads through a single system.

Its architecture contains approximately 1 trillion parameters, but only 49 billion parameters are active during an operation. This distinction allows Mistral to describe the model as extremely large while using a smaller active portion for individual tasks.

The preview can be accessed through Mistral Studio. The company is also preparing an open-weight release that could give developers and businesses greater control over deployment.

Mistral Large 4 at a glance

Why Is Mistral Large 4 Important?

Large 4 is notable because Mistral is targeting several advanced AI applications rather than focusing only on chat and text generation.

The model is being evaluated for:

• Software development
• Cybersecurity research
• Business automation
• Scientific problem-solving
• Visual understanding
• Technical document analysis
• Tool-based workflows
• Multilingual applications

Its planned open-weight release is another important factor. Businesses could potentially deploy the model in private cloud or on-premise environments instead of relying only on a hosted AI service.

Mistral Large 4 for Cybersecurity

Cybersecurity is one of the major areas highlighted by Mistral.

The company says it is conducting real-world red-team testing with cybersecurity leaders, selected partners, and government authorities.

Mistral reports that Large 4 ranked among the top five models in the Artificial Analysis Cyber Index. It also reports an 82% score on a test involving the reproduction and patching of a real vulnerability in open-source software.

On Cybench’s 40 security exercises, Mistral reports a 93% success rate.

The company says internal testing also showed potential for:

• Malware analysis
• Vulnerability prioritisation
• Detection-rule creation
• Security research
• Incident-response support

These are company-reported benchmark results and should not be treated as proof that the model can handle every cybersecurity situation reliably.

Mistral Large 4 Coding Performance

Software engineering is another major focus of the new model.

Mistral reports the following results from Artificial Analysis coding evaluations:

• 61.7% on DeepSWE v1.1

• 59.4% on SWE-Atlas-QnA

• 28.3% on Terminal-Bench 4

• 49.8% Coding Agent Index score

A separate human evaluation by Surge AI asked professional annotators to score coding outputs without knowing which model produced them.

Large 4 Preview received a 3.74/5 rating, placing second among five models tested. Claude Opus 5 received a 4.22 score in the same evaluation.

These results indicate that Mistral is positioning Large 4 for complex development tasks rather than basic code generation alone.

Can Mistral Large 4 Automate Business Workflows?

Mistral is also testing the model for agent-style business automation.

On AutomationBench, which evaluates 657 workflows involving applications such as Gmail, Google Sheets, Slack, and Salesforce, Mistral reports a 59.9% score.

These tasks require AI systems to understand instructions, interact with applications, perform multiple steps, and complete a requested workflow.

This capability could make Large 4 useful for enterprise assistants, workflow automation, research agents, and software tools that need to interact with external services.

Mistral Large 4’s Multimodal Capabilities

Large 4 is designed to process more than text.

Mistral has demonstrated the model with:

• Technical drawings
• PDF documents
• Satellite imagery
• Visual-grounding tasks

The company reports a 42% result on Dense 200 visual grounding, compared with 41% for GPT-6-Astra in its testing.

Visual grounding allows an AI system to connect information in an image with specific objects, regions, or details. This can be useful for technical documents, engineering material, research, and geographic imagery.

How Was Mistral Large 4 Trained?

Mistral says Large 4 was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs in its European datacentres.

Its training data covers more than 160 languages, including all official languages of the European Union.

The company has also described the model with the nickname “le Chonk,” which began as a meme during development.

Mistral says the model uses reinforcement learning across several training environments. These include conversational tasks, scientific reasoning, factuality, safety alignment, and long-running tool-use scenarios.

The training setup can use resources such as code sandboxes, web search, and external APIs. Results can then be checked using reward models, unit tests, LLM judges, and static checks.

Mistral Large 4 Open-Weight Release

Mistral plans to release the Large 4 model weights by the end of October 2026.

The company says the release will also provide additional architecture information, benchmark results, and details about its post-training methods.

An open-weight version could be particularly valuable for organisations that need greater control over AI deployment.

Private-cloud and on-premise operation may help companies manage sensitive data, internal security requirements, infrastructure, and access policies more directly.

What Can Mistral Large 4 Be Used For?

Potential applications include:

• Software Development: Developers can use the model for coding, debugging, testing, code analysis, and agent-based programming workflows.
• Cybersecurity: Security teams could apply it to vulnerability research, malware analysis, and defensive security tasks.
• Business Automation: The model can support workflows involving business applications and multi-step tasks.
• Document Analysis: Its multimodal capabilities can help process information contained in PDFs and other visual documents.
• Research: Mistral’s training approach includes scientific problem-solving, making the model relevant to research-oriented applications.
• Visual Analysis: Technical drawings, satellite images, and other visual content can be processed as part of multimodal workflows.

Conclusion

Mistral Large 4 is a significant addition to the growing market for large multimodal AI systems. Its 1 trillion-parameter architecture, coding capabilities, cybersecurity testing, business automation features, and planned open-weight release make it particularly relevant to developers and enterprises.

The public preview provides an early look at the model, but the October 2026 weight release could be the more important milestone. Independent testing and real-world deployments will ultimately show how Large 4 compares with competing AI systems.

For now, Mistral’s reported benchmarks suggest that the company is aiming to make Large 4 a flexible AI system for software development, security, automation, research, and multimodal workloads.

Frequently Asked Question

What is Mistral Large 4?

Mistral Large 4 is a multimodal AI model with 1 trillion total and 49 billion active parameters.

Is Mistral Large 4 available?

Yes. Its public preview is available through Mistral Studio, with weights planned for release by the end of October 2026.

What can Mistral Large 4 do?

It supports coding, cybersecurity, automation, research, document analysis, and visual tasks.

How many parameters does it have?

Mistral Large 4 has 1 trillion total parameters and 49 billion active parameters.