Mistral Large 4, or “Le Chonk,” is an open-weight multimodal model with 1 trillion parameters and 52 billion active parameters. Its strengths lie in cybersecurity, coding, agentic workflows, and understanding multiple data formats.
It is now available for testing through Mistral Studio, while the weights are scheduled for release by the end of the month.
What Was Launched This Time
The source does not provide product imagery or design details for Mistral Large 4, so a conceptual illustration is used instead. The model is currently in public preview API status and available for testing through Mistral Studio.
The weights will be released by the end of the month. Anyone who wants to test its capabilities before using it in production can start with the API now.
When System Defense Requires Doing What General Models Refuse
Security teams sometimes need to prove that a vulnerability exists, analyze malware, or write detection rules. However, systems with provider-enforced refusals can obstruct critical work during an incident.
Mistral Large 4 takes an approach based on open weights and self-deployment, allowing organizations to run the model on a private cloud or on-premise under their own policies. This provides greater control and oversight for security work without relying on the provider’s decisions at every step.
Where Mistral Large 4 Fits in the Mistral Family
Mistral Large 4, or ML4, is Mistral’s largest and most capable model at launch. The company positions it as the foundation for future specialized Mistral models and models optimized for different tasks.
What makes it particularly interesting is that ML4 uses the same environment as Mistral Forge for training, fine-tuning, and reinforcement learning. This makes Forge similar to a toolkit for directly adapting models to each organization’s specific needs.
What Changed from the Previous Model to Le Chonk
The source does not specify the previous model’s name or provide direct comparison figures, so this table should be read as a comparison between ML4 and undisclosed information about the previous model.
| Factor | Mistral Large 4 | Previous model |
|---|---|---|
| Launch status | Public preview | Not specified |
| Model size | 1 trillion parameters; active 52 billion | Not specified |
| Core capabilities | Coding, agentic workflows, cybersecurity, and enterprise workloads | Not specified |
| Multimodal support | Natively multimodal model | Not specified |
| Weight release | To be released by the end of the month | Not specified |
| Infrastructure | Trained with 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data center | Not specified |
Capabilities That Become Clear in Real-World Use
Security teams can use ML4 to reproduce vulnerabilities in open-source software and write patches to fix them. In one test from the Artificial Analysis Cyber Index, it scored 82%. This type of work helps prove that a vulnerability exists before mitigation begins.
Developers can have the model understand a repository and handle complex tasks through the terminal. This is well suited to projects that require reviewing multiple parts of a codebase simultaneously.
For business workloads, ML4 agents can gather information, use tools, and deliver work according to workflows through apps such as Gmail, Google Sheets, Slack, and Salesforce.
For longer-term tasks, the model can work with spreadsheets, slides, and PDFs, making it suitable for teams that need to turn large amounts of data into ready-to-use documents.
How It Ranks Against Other Models
ML4 Preview sits among the top performers based on the available test results, particularly for coding agents, agentic tasks, and business workloads. However, many competitors only report that ML4 scored higher, so the results should not be interpreted as showing a specific margin.
| Factor | ML4 Preview | DeepSeek V4 Pro |
|---|---|---|
| Coding tasks | 49.8% Coding Agent Index | ML4 scored higher than DeepSeek V4 Pro 0813 |
| Agentic tasks | 59.9% AutomationBench | ML4 scored higher |
| Business tasks | 1,393 Elo on AA-Briefcase | ML4 scored higher |
| Cybersecurity tasks | 82% vulnerability reproduction and remediation test | No score specified |
Clear Strengths and Limitations to Know
Mistral Large 4 stands out in cybersecurity, with open weights and support for private cloud or on-premise deployment. This makes it suitable for organizations that want to control their own data and policies. It also supports more than 160 languages and achieves strong coding results, such as 61.7% on DeepSWE v1.1.
Pros
- +Advanced cybersecurity capabilities
- +Supports open weights, private cloud, and on-premise deployment
- +Supports more than 160 languages
- +Performs well in coding and agentic workflows
Cons
- −Still in public preview
- −Architecture and post-training details have not been disclosed
- −Pricing information is not yet available
- −More benchmark results have yet to be released
Actual Costs That the Source Has Not Yet Disclosed
The source does not yet specify API pricing, licensing fees, or infrastructure costs, so it is not possible to estimate the cost numerically. Organizations must first decide whether to use the preview API or wait for the weights in order to control the system themselves.
Those choosing self-deployment must consider the costs of private cloud or on-premise infrastructure, as well as the operational burden and governance required to comply with organizational policies. Costs and responsibilities will differ between using infrastructure managed by Mistral and running the entire system independently.
Who Should Try Mistral Large 4 Right Now
Cybersecurity teams, software development teams working with complex repositories, and organizations seeking agents to assist with business workflows are good candidates for trying the public preview API now. This is especially true for those interested in open weights and deployment under their own policies.
Anyone seeking pricing, stability, or more complete architectural details should wait for additional information and the release of the weights.
Made for
- Cybersecurity teams
- Software development teams and complex repositories
- Organizations seeking agents for business workflows
- Organizations that prioritize open weights and control over their own deployment
Think twice
- Those who need complete pricing, stability, and architectural details
Skip this one
- Those who need complete information before getting started — wait for additional details and the release of the weights
The Key Decision Is Not Just the Score, but Control Over the Model
Mistral Large 4 is interesting because it combines advanced capabilities with open weights and self-deployment. Organizations therefore have the freedom to choose how to deploy and use the model under their own policies, making it suitable for work that requires both capability and autonomy.
The preview API is now available through Mistral Studio. Further details on the architecture, additional benchmarks, and post-training methodology will be disclosed later. Anyone evaluating the model for real-world work should follow this information as part of the decision-making process.