Mistral releases open weights for its 1 trillion-parameter ML4 model

A new standard for open weights

Mistral has officially released the weights for its Large 4 model, which the company nicknamed "Le Chonk." This release follows a three-week preview period where the model was refined using reinforcement learning guided by direct customer feedback. The goal of these updates was to sharpen the model's performance in specific high-value areas like coding and cyber defense.

The model is capable of processing both text and images to produce text and code outputs. While it is a massive 1 trillion-parameter model, it is designed to be efficient by only activating 52 billion of those parameters for any single task. This architecture allows it to handle complex requests without requiring the full weight of the model to be active at all times.

As the weights are now available on Hugging Face, the model can be downloaded and run by various developers. Mistral is also providing a quantized version, which is a compressed format that allows the model to run on a smaller hardware footprint of four Nvidia B200 or B300 GPUs. This makes the model more accessible to organizations that do not have massive data centers.

Coding and security gains

The most visible improvements in the new release are found in agentic coding, which refers to AI that can perform multi-step tasks in a terminal. On the Terminal-Bench 4.0 test, ML4 improved its score from 28% during the preview to 41%. This puts it slightly ahead of the GLM-5.3 model, which scored 40% on the same benchmark.

In addition to raw performance, the model shows significant efficiency gains. According to the Artificial Analysis Coding Agent Index, ML4 rivals GLM-5.3 while using 47% fewer tokens per task on average. This reduction in token usage is important for coding agents because it saves significant costs and time when these agents are running for several hours at a time.

The model also shows strong results in cyber defense tasks. It scored 79% on CyberGym-E2E, a test designed to see if a model can find, reproduce, and patch real memory-safety bugs in open-source C and C++ projects. This score matches the MiMo-V2.6-Pro model and beats the DeepSeek V4.1 Flash model, which scored 76%.

Enterprise-focused features

During the preview phase, early testers—which Mistral identified as governments and enterprises—requested more control over the model's reasoning time. In response, the company added low, medium, and high reasoning modes to the model. The high reasoning mode was specifically optimized to be 50% more efficient than it was during the initial preview.

These features are already being utilized by customers for automated penetration testing. In these scenarios, specialized agents use the model to probe applications for security vulnerabilities and then write up a report on their findings. This allows companies to proactively identify and fix weaknesses in their own software before they can be exploited.

Mistral's VP of Science, Pierre Stock, noted that the company wanted to provide these options to ensure users have the right balance of speed and depth. By offering different reasoning modes, the model can be tuned for quick answers or for deep, methodical problem-solving depending on the specific needs of the user.

Efficiency and resource scale

Mistral claims to have built ML4 with a fraction of the resources used by its competitors. While the exact costs of Chinese labs are often undisclosed, Mistral highlights its own training data. The model was trained on 3,800 Nvidia Grace Blackwell GPUs located in Europe, and the reinforcement learning was conducted on its own compute in France.

The company emphasizes that the reinforcement learning process drew about 10 megawatts of power. CEO Arthur Mensch pointed out that while other labs have raised hundreds of billions of dollars to develop similar capabilities, Mistral raised roughly $6 billion over the last three years. This suggests a high level of efficiency in how they train and deploy their models compared to larger, more heavily funded rivals.

The company is also planning to release smaller, distilled versions of the model in the coming months. These smaller versions will be more accessible to individual users, with one version expected to run directly on a laptop. While Mistral has not provided specific dates for these releases, they aim to make the technology usable across a wider variety of devices.

The philosophy of open weights

The decision to release the weights of ML4 is rooted in Mistral's belief that AI should be as accessible as electricity. Pierre Stock argued that no company should have to depend on a provider who can cut off access to a model from one day to another. By releasing the weights, Mistral ensures that once the model is online, the access cannot be withdrawn by the company.

This stance contrasts with some rivals who argue that open models carry more risk. However, Stock applied this same logic to the Chinese labs that dominate the open model space. He suggested that nothing guarantees those labs will continue to publish weights, and their models could potentially be turned into attack vectors with less transparency.

Moving forward, Mistral aims to create a consistent cadence of frontier-level releases. The company wants to move beyond reaching the frontier once and instead establish a regular rhythm of innovation. For now, ML4 is available through the Mistral API, on Hugging Face, and to some users of the Vibe platform.