Mistral releases weights for its 1-trillion-parameter Large 4 model

The release of Le Chonk

Mistral has officially released the downloadable weights for its new 1-trillion-parameter model, Mistral Large 4. The company nicknamed the model Le Chonk and made it available on Hugging Face and through its own API. This release comes three weeks after the company provided an initial preview of the model's capabilities.

By releasing the weights, Mistral is allowing companies to download the model's 'brain' to run on their own private servers. This is a significant move for privacy-conscious organizations because it ensures that sensitive data never has to leave the internal network. It provides an alternative to using third-party APIs where data is processed on someone else's infrastructure.

The model was developed using three additional weeks of reinforcement learning following its initial preview. According to Mistral, these refinements helped the model achieve high performance on coding tasks. Specifically, it reportedly ties with GLM-5.3 on a coding-agent index while being more efficient with its token usage.

How the architecture works

Despite having 1 trillion parameters, Mistral Large 4 utilizes a mixture-of-experts architecture. This means that only about 49B of its parameters are active at any given time. The system functions by activating specific 'specialists' within the model to handle different parts of a request rather than running the entire network at once.

The model is designed to be versatile, supporting both text and image inputs. It also offers three different reasoning settings to help users tune its behavior for specific tasks. For long-form content or complex data analysis, it supports a 1M-token context window, which allows it to process very large amounts of information in a single prompt.

From a hardware perspective, running the model is a demanding task. Mistral recommends a setup that splits the model across eight GPUs to handle the workload effectively. While downloading the weights is relatively inexpensive, the actual compute power required to run a model of this scale remains a significant barrier for smaller entities.

Security and restricted access

Mistral has taken a tiered approach to the release of Large 4 due to concerns about its capabilities in cyberattacks. The company noted that its testing suggested the model could potentially increase the real-world abilities of attackers. However, the model did refuse 96.6% of malicious cyber requests on a specific set of benchmarks.

Because of these risks, Mistral split the release into two versions. The general model is available for the public to download, but a stronger cyber variant is restricted to vetted partners like Orange Cyberdefense. This restricted version is intended specifically for penetration testing and red-teaming, with contracts limiting its use to systems that the testers own or are authorized to probe.

This strategy mirrors moves by other major AI labs, such as Google and Anthropic, which have also limited certain cyber-focused capabilities to selected defenders. Mistral's approach aims to provide high-end security tools to professionals while attempting to prevent the general public from accessing the most potent offensive capabilities.

Licensing and industry impact

The licensing for Mistral Large 4 is slightly more restrictive than previous versions. It uses a modified MIT license that allows for internal self-hosting by a company. However, any entity that intends to sell the model as a service must obtain a separate commercial license, which is a tighter restriction than the Apache 2.0 license used for Large 3.

This shift in licensing reflects the growing commercial value of frontier models. It allows Mistral to maintain a level of control over how the model is monetized while still providing the freedom for internal use. This is particularly important for large enterprises that want to integrate AI into their internal workflows without creating a public-facing product.

Several major companies, including Cisco and Cloudflare, have already expressed support or interest in testing the model. For industries like banking and government, the ability to host a model of this scale locally is a major milestone. It allows them to explore frontier-level AI reasoning while adhering to strict data residency and security requirements.

The road ahead for Mistral

Mistral is positioning itself as a leader in providing high-performance, controllable AI options. By offering a model that can be hosted on-premise, they are betting that corporate control is a primary selling point. The company is also exploring a European-centric option involving a €3B round led by Samsung to train models on Nvidia GPUs in European data centers.

While Microsoft has already committed billions to use part of that capacity, the focus remains on providing diverse infrastructure choices. This allows users to choose where their models run without necessarily having to cut ties with U.S. companies. It creates a landscape where geographical and infrastructure preferences can dictate AI adoption.

The next major step for the community will be the independent replication of Mistral's own benchmarks. While the company reports impressive results on coding and reasoning, the broader community will need to verify these numbers. The full technical report will be the key document for researchers to understand the exact limits and safety guardrails of the model.