TL;DR: ML4 is a multimodal open-weight model (supporting text, images, and documents) that stands out in cybersecurity, coding, and agentic tasks. It is suitable for organizations that want control over AI usage and want to run it on their own systems. If it were up to me, I would recommend starting with API experimentation, then evaluating self-deployment once the weights are available.
Mistral Large 4 (ML4) is a multimodal open-weight model (supporting text, images, and documents) with 1 trillion total parameters and 49 billion active parameters. Its strengths cover cybersecurity, coding, agents, and knowledge work such as spreadsheets, slides, and PDFs.
ML4 is currently available as a public preview through Mistral Studio. The model weights, or parameter set, are scheduled to be released by the end of the month. It is suitable for organizations that want to control their AI usage and run it on a private cloud or their own infrastructure.
Mistral Large 4 at a Glance
Mistral Large 4 is a 1 trillion-parameter multimodal model with 49 billion active parameters, and its weights are set to be released for open-weight use.
From a product lineup perspective, ML4 is currently Mistral’s flagship model. Smaller models are better suited to tasks requiring lower costs and fewer resources.
When Security Teams Need to Prove Vulnerabilities Without Being Refused
Vulnerability research and incident response can be disrupted when provider-level refusals block requests that have legitimate security purposes. During an active incident, losing capabilities midway becomes an additional layer of risk.
ML4 is designed for this type of work. It supports vulnerability exploitation, malware analysis, vulnerability prioritization, and detection-rule writing. In one test, ML4 scored 82% and solved 93% of the challenges in Cybench.
The key point is that organizations can use the open weights and self-deployment—that is, install and manage the model on their own systems—on a private cloud or on-premise infrastructure. This allows them to control usage in accordance with their own security policies.
Where ML4 Fits in Mistral’s Direction
At the time of its announcement, ML4 was Mistral’s largest and most capable model. The company positioned it as the foundation for specialized Mistral models and models optimized for different tasks.
Releasing the weights and supporting self-deployment allows organizations to control their own AI usage, from setting policies to running the model in their own environments. This direction is well suited to critical workloads that require both capability and autonomy.
From Earlier Models to ML4: What the Sources Confirm
| Factor | Earlier model | Mistral Large 4 |
|---|---|---|
| Model size | No data from primary sources | 1 trillion parameters; active 49 billion |
| Capabilities | No data from primary sources | natively multimodal |
| Status | No data from primary sources | public preview through Mistral Studio |
| Weights release plan | No data from primary sources | By the end of the month |
The primary sources do not provide numerical details about earlier models for comparison. We can therefore conclude only that ML4 is currently Mistral’s largest and most capable model, with access available through a preview API.
How Its Capabilities Translate into Real-World Work
Security teams can use ML4 to simulate vulnerabilities in open-source software and help write patches. It scored 82% in one test and solved 93% of the Cybench challenges, making it suitable for teams that need to investigate and respond to real incidents.
For development teams, ML4 understands repositories and can work through complex terminal tasks. It scored 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.
In business workflows, ML4’s agents can use tools and work across Gmail, Google Sheets, Slack, or Salesforce. It scored 59.9% on AutomationBench.
When work needs to be delivered end to end, ML4 can also create spreadsheets, slides, and PDFs. Its score on AA-Briefcase was 1,393 Elo, making it suitable for knowledge work that involves multiple steps and results in ready-to-use documents.
ML4 ranks near the top across multiple evaluations, but it does not lead in every category. In a blind coding evaluation, ML4 Preview ranked 2nd out of 5 models, while Claude Opus 5 ranked 1st.
| Factor | ML4 Preview | Claude Opus 5 | Kimi K3 | GLM-5.3 | GLM-5.2 | DeepSeek V4 Pro |
|---|---|---|---|---|---|---|
| Blind coding evaluation | Ranked 2nd, 3.74/5 | Ranked 1st, 4.22/5 | 3.59/5 | 3.60/5 | 3.40/5 | Not specified |
| AutomationBench | 59.9% | Not specified | Lower than ML4 | Not specified | Not specified | Lower than ML4 |
| AA-Briefcase | 1,393 Elo | Not specified | Not specified | Not specified | Not specified | Lower than ML4 |
Overall, ML4 stands out in coding and agentic tasks that require continuous work, while Claude Opus 5 still leads in blind coding quality scores.
Clear Strengths, but Limitations Remain to Be Seen
ML4 stands out in cybersecurity, scoring 82% in a vulnerability test and 93% in Cybench. Another strength is its support for more than 160 languages, making it suitable for organizations operating across multiple countries.
Pros
- +High cybersecurity scores among open-weight models
- +Supports more than 160 languages
- +Can run on a private cloud or on-premise
Cons
- −Still in public preview
- −Architecture details have not yet been disclosed
- −Post-training methodology has not yet been disclosed
For now, it is best suited to API experimentation. For production deployment, it is advisable to wait for the weights and additional details, which are expected to be released by the end of the month.
Costs to Consider Beyond API Usage
Using the preview API makes it possible to start experimenting immediately, with Mistral managing the infrastructure. However, using the open weights for self-deployment on a private cloud or on-premise infrastructure shifts the burden of system and management costs…