ML4 AI Model Unveiled by Mistral with Customizable Control and Data Sovereignty

French artificial intelligence (AI) lab Mistral has unveiled its latest model, ML4, which is being touted as a comprehensive solution for all business defense needs. The new model, nicknamed "Le Chonk" due to its massive trillion-parameter size, offers customizable control and data sovereignty, setting it apart from proprietary models.

In recent times, the AI security landscape has been marred by incidents that have raised concerns about the safety of both open and proprietary models. Initially viewed as riskier due to their malleability, open models are now being considered as a viable option for cyber defense, especially in light of the vulnerabilities discovered in proprietary models from Anthropic and OpenAI.

Mistral's ML4 model is designed with security at its core, allowing users to maintain control over their defenses. This stands in contrast to proprietary models, which can be subject to access revocation by frontier labs at any time. The company's co-founder, Guillaume Lample, emphasized the importance of user control in AI security.

The latest developments come on the heels of a summer marked by numerous high-profile AI model hacks that highlighted the need for reliable and customizable security solutions. ML4 prioritizes cyber defense capabilities, enabling enterprises and governments to defend themselves against threat actors who exploit vulnerabilities in closed models.

Mistral's focus on security is driven by its enterprise clients' requirements, reflecting an industry trend that emphasizes the importance of robust defenses. The company's participation in Nvidia's Open Secure AI Alliance, a cross-industry partnership aimed at promoting open models and democratizing AI security, underscores its commitment to this cause.

In a briefing with Mistral's leadership, including co-founder Guillaume Lample and VP of Science Pierre Stock, it became clear that user control and data sovereignty are key considerations in the development of ML4. The model's customizable nature is designed to address the evolving needs of businesses and governments seeking to protect themselves against increasingly sophisticated cyber threats.

Mistral's latest model, Le Chonk, is being touted as a game-changer in AI cybersecurity, offering businesses and governments a customizable solution that can be tailored to their specific needs.

The company claims that its model, ML4, outperforms the best models from other leading developers, including Kimi, Deepseek, and Meta. This assertion has been made without the need for vendor lock-in, allowing enterprises to fully own and control their cyber defense capabilities.

In a move that highlights the importance of open-source AI security tools, Mistral is releasing Le Chonk in public preview, with the model weights set to be released on October 27. This will give developers, cybersecurity leaders, and state authorities a month to assess ML4's capabilities and behavior in real-world environments.

As part of its testing process, Mistral has partnered with initial testing partners who will gain access to an expanded version of ML4 with enhanced cybersecurity capabilities. This is similar to the approach taken by other proprietary American labs, such as OpenAI and Anthropic, which have been releasing less-guardrailed versions of their models for testing purposes.

Early analysis suggests that Le Chonk excels in finance and multimodal use cases, competing on par with pricier proprietary models like GPT-6 Astra. It has also shown impressive results in computer vision tasks, meeting or slightly outperforming DeepSeek models.

Outside of security, the model's capabilities are being put to the test in various benchmarks, including Harvey's Legal Agent benchmark, where it achieved a new high of 15% for open-weight models.

The benchmark scores achieved by Mistral's new Le Chonk model should be viewed with caution due to the prevalence of models that cheat in these evaluations. The controversy surrounding cheating AI models is a well-documented issue, with some labs accused of distilling proprietary models from others to gain an advantage.

Mistral has emphasized its commitment to integrity and innovation, stating that it does not participate in the practice of distilling or ripping off proprietary models from other labs. This stance is reflected in the company's statement that it operates independently and does not draw inspiration from rival models.

In addition to its focus on security, Mistral has also highlighted its dedication to sovereignty, a key concern for many organizations, particularly those in Europe. The company has promised customers greater flexibility in deploying the Le Chonk model, allowing them to choose between self-deployment or accessing it through the company's API from their preferred region.

Mistral's own resources and infrastructure have been put to the test in training its competitive Le Chonk model. According to the company, ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, using Mistral's own data centers in Europe.

Mistral's Le Chonk model is set to improve its performance over time, as the company anticipates significant enhancements in the next few months.

The new AI cybersecurity model will serve as a foundation for future specialized and optimized models developed by Mistral, expanding the company's offerings in this area.