Short summary: Mistral Large 4, or ML4, is a 1-trillion-parameter multimodal model from France. Its key feature is a plan to release its weights after safety testing, but for now it is available only through a public guardrail endpoint.
Mistral Large 4 (ML4), or “Le Chonk,” is a 1-trillion-parameter multimodal model from France. Mistral positions it as a European alternative between American closed models and open models that often come from China.
For now, it is available only through a public guardrail endpoint. Mistral plans to release the weights after completing safety testing in three weeks. The company is focusing on cybersecurity, finance, and chip design.
What ML4 Is at a Glance
Mistral Large 4, or ML4, is a 1-trillion-parameter multimodal model. Its position is as an option between closed and open models.
This image summarizes ML4 simply: a large model at the center that accepts multiple types of tasks, with Mistral attempting to compete on both sides.
The Problem for Organizations That Want More AI Oversight
Organizations and institutions are placing greater importance on AI model safety, especially Mistral’s core customers, which include enterprises and institutions. They want options whose operations can be inspected more thoroughly than those of closed models.
What makes ML4 interesting is that Mistral says an open-weight model can be inspected more easily. However, it is currently available only through a public guardrail endpoint. The company plans to release the weights after completing safety testing in three weeks, meaning organizations must wait to assess both its safety and practical usage before deciding.
Where ML4 Fits in Mistral’s Direction
ML4 is a step Mistral is using to reaffirm that it remains a frontier lab, not merely a model provider. It positions itself as AI’s “third way,” as described by French President Macron, between closed models from the United States and open models that often come from China.
Put simply, closed models are like an older sibling that controls everything itself, while ML4 is trying to position itself as an option that allows greater inspection and has its own model behind it.
Although Mistral previously chose to host Chinese models, the company says this does not mean it is becoming merely an inference provider. Having Le Chonk, or ML4, with 1 trillion parameters is therefore evidence that Mistral still wants to develop frontier models itself.
From Hosting Chinese Models to Le Chonk
| Factor | Before the ML4 launch | During the ML4 launch |
|---|---|---|
| Company role | Affirms that it is not merely an inference provider | Positions itself as a frontier lab |
| Model approach | Hosts Chinese models | Develops ML4 with 1 trillion parameters |
| Open-weight position | Model weights are not yet open | Not released immediately; planned release after safety testing |
| Enterprise customers | Prioritizes enterprises and institutions | Focuses on cybersecurity, finance, and chip design |
Overall, Mistral is attempting to move from being a model provider to developing frontier models itself. However, ML4 still needs to undergo safety testing before its weights can be released for full open-weight use.
Features That May Matter in Real-World Use
ML4 is a multimodal model, so it may be suitable for tasks that require reading and connecting multiple types of information, such as images and text. However, Mistral has not yet provided benchmark results to confirm this.
Its size of 1 trillion parameters raises expectations for performance, especially on specialized tasks. However, this remains a target rather than a confirmed test result.
Preparing to release the weights after safety testing could help organizations inspect the model and govern its use in greater detail. Its actual readiness, however, still depends on the release of the weights.
Mistral says it trained ML4 using 4,000 of its own Nvidia GPUs, reflecting an approach based on controlled computing resources. All of this represents potential use cases based on Mistral’s stated goals, not conclusions from actual testing.
ML4 Compared with the Options Organizations Are Considering
ML4 is still awaiting benchmarks and the release of its weights after approximately three more weeks of safety testing. It may suit organizations that want an inspectable model in the future, but they should wait until it is truly ready before making a decision.
In terms of the product lineup, ML4 is not yet the “older sibling” that has proven itself, nor is it the “younger sibling” ready for everyone to experiment with fully openly. If I were advising an organization, I would recommend following the benchmark results and waiting for the weights to be released before using it for critical work.
| Factor | ML4 | Closed models from the United States | Open models often developed in China | Other open-weight models outside China |
|---|---|---|---|---|
| Access to weights | Not yet released; available through a guardrail endpoint | Closed | Open | Information unavailable |
| Benchmark status | Results pending | Information unavailable | Information unavailable | Information unavailable |
| Inspectability | Likely to be inspectable once the weights are released | Limited inspectability | More inspectable than closed models | Information unavailable |
| Safety | Undergoing safety testing | Closed and controlled by the provider | Usage risks must be managed | Information unavailable |
| Focus areas | Cybersecurity, finance, chip design | Information unavailable | Information unavailable | Open-weight models outside China |
Strengths and Limitations Evident from the Current Information
ML4 stands out for its 1 trillion parameters, multimodal capabilities, and training on Mistral’s own compute, using 4,000 Nvidia GPUs, along with a plan to release its weights after safety testing is complete.
Pros
- +1 trillion parameters
- +Supports multimodal tasks and focuses on cybersecurity, finance, and chip design
- +Trained using Mistral’s own compute
- +Plans to release the weights in three weeks
Cons
- −Benchmark results have not yet been disclosed
- −Currently accessible only through a public guardrail endpoint
- −Safety testing must be completed before the weights are released
The Still-Unknown Cost of Deploying ML4
ML4 was trained using Mistral’s own compute, with Nvidia GPUs.