Mistral releases open-weight ML4 model for coding and cybersecurity

A new flagship for open AI

Mistral has officially released the weights for its latest open model, Mistral Large 4, also known as ML4. The company describes the system as its flagship open frontier model, designed to compete with the largest AI systems currently available. It is a natively multimodal system, meaning it can process different types of data, and it features 1 trillion parameters with 52 billion active parameters.

The release follows a period of intensive testing with cybersecurity partners and various customers. Mistral spent three weeks performing additional reinforcement learning training after an initial preview of the model. This process helped refine the model's capabilities before it was made available to the public on the Hugging Face platform.

The company is positioning ML4 specifically for high-stakes environments like coding and cybersecurity. By providing the open weights, Mistral enables organizations to host the model on their own internal systems. This is a strategic move for companies that require strict control over their data and software access.

Efficiency and reasoning modes

One of the primary technical highlights of ML4 is its efficiency compared to other large models. On the Artificial Analysis Coding Agent Index, the model uses 47% fewer tokens per task than the GLM-5.3 model while producing similar results. This means the model can complete tasks using less computational 'language' units, which can reduce costs and improve speed.

Mistral also introduced three distinct reasoning modes for the model: low, medium, and high. These settings allow users to manually adjust the level of thinking the AI applies to a specific task. This flexibility helps users balance the need for deep logic against the desire for faster, simpler responses during different workflows.

The model's performance on technical benchmarks has also seen a significant jump. Mistral reported that ML4 climbed from a 28% score to 41% on Terminal-Bench 4.0. This puts the model on par with GLM-5.3 on that specific benchmark, which measures how well an AI can interact with a command-line interface.

Private infrastructure for defense

For government bodies and large enterprises, the ability to deploy ML4 on private infrastructure is a major selling point. Because the weights are open, these organizations can use the model to detect and analyze threats across their IT systems. This setup allows them to respond to cyberattacks without having to grant a third-party provider access to their internal networks.

Mistral notes that this approach is particularly appealing to regulated sectors that must adhere to strict data privacy laws. By keeping the model on-site, these organizations can maintain tighter control over their proprietary code. It removes the risks associated with sending sensitive information to a cloud-based AI service for processing.

The model was trained using 3,800 Nvidia Grace Blackwell graphics processing units located in European data centers. Mistral claims this training path allowed them to achieve state-of-the-art results with lower capital demands than the developers of closed-source models. The company aims to prove that high-quality science can be delivered more efficiently than by spending hundreds of billions of dollars.

Industrial and multimodal uses

Beyond cybersecurity and coding, Mistral has outlined several diverse industrial applications for ML4. In the manufacturing sector, the model can be used to convert two-dimensional blueprints into three-dimensional models within computer-aided design software. This could streamline the transition from initial design to physical production.

Financial institutions may find use for the model in large-scale data analysis and information extraction from various filings. It can also assist with complex spreadsheet work and other data-heavy tasks. These applications benefit from the model's ability to handle large amounts of structured and unstructured information efficiently.

The multimodal nature of ML4 also opens doors for geospatial tasks involving satellite and aerial imagery. Mistral suggests the model could be used for wildfire detection, monitoring crop health, and identifying risks to power line infrastructure. These use cases demonstrate the model's versatility across different types of visual and textual data.

The open-weight philosophy

The launch of ML4 occurs during a heated debate over whether leading AI systems should be open-weight or kept behind proprietary interfaces. Open-weight models allow developers to download the parameters and run them on their own hardware. While this requires significant computing resources to run at scale, it prevents the concentration of AI power in the hands of a few companies.

Mistral CEO Arthur Mensch has stated that open source is essential to ensure that state-of-the-art AI does not end up concentrated in the hands of a few actors. He contrasted Mistral's funding, which reached roughly $6 billion in three years, with the hundreds of billions raised by some rivals. The company argues that their approach focuses on delivering quality science to everyone.

ML4 is now available through Mistral's application programming interface as well as the Hugging Face release. This follows the recent introduction of Voxtral 3, a collection of voice models and agents. Together, these releases show Mistral's effort to expand its lineup across different formats while maintaining its focus on open-weight accessibility.