Mistral Launches ML4: A Game-Changer in Open AI Models

On October 6, French AI lab Mistral launched the Mistral Large 4 (ML4), affectionately termed "Le Chonk" due to its impressive trillion parameters. This new model is designed to raise the bar for open models, particularly in the realms of coding performance and cybersecurity.

Significant Improvements in Coding Efficiency

The ML4 model has shown notable advancements in its coding abilities, reflected in its updated score on the Terminal-Bench 4.0. Initially, the model scored 28%, but thorough optimization has led to a jump to 41%. This newfound score places ML4 ahead of Opencode's GLM-5.3 by a slim margin, although it remains just one point shy of GLM-5.3 in the overall performance index, matching Codex GPT-6.1 Terra with a score of 48.

It's crucial to approach benchmark scores with caution, as many factors can influence the outcomes and some models may not adhere to testing protocols. Nevertheless, the performance leap in ML4 is notable considering the model underwent just a few weeks of additional training. Mistral claims that ML4 accomplishes tasks with 47% fewer tokens per job compared to its competitor GLM-5.3, making it an economical option without performance trade-offs.

Enhanced Cybersecurity Features

Mistral has also emphasized the importance of cybersecurity in its ML4 model. The company focuses on providing customizable controls and ensuring data sovereignty, a necessity increasingly demanded by organizations worldwide. Mistral has partnered with various developers to thoroughly evaluate the model's cybersecurity prowess, similar to efforts seen in proprietary labs like OpenAI and Google.

In initial tests, ML4 demonstrated potential in cybersecurity, achieving favorable results on the CyberGym benchmark by outperforming DeepSeek V4.1 Flash. This benchmark evaluates models based on their capability to identify and rectify memory-safety vulnerabilities in C/C++ open-source projects, effectively patching them while maintaining project integrity.

The environment around AI security is continuously evolving, and Mistral's strategy positions open models like ML4 as promising alternatives to proprietary models, which have struggled with their own security threats. Mistral's co-founder, Guillaume Lample, articulated that ML4 was specifically built to empower users to maintain control, in contrast to proprietary models that can restrict access.

Optimizing Resources Despite Limitations

Developing a state-of-the-art model amidst a compute shortage poses substantial challenges, particularly for a smaller lab like Mistral. The ML4 model was trained from the ground up using 4,000 Nvidia Grace Blackwell GPUs over two months, hosted within Mistral's data centers in Europe. For perspective, larger organizations like OpenAI utilized approximately 100,000 GPUs for their own models.

As the AI landscape continues to shift, Mistral's introduction of ML4 serves as a significant marker in the competitive space between open and proprietary models. With its impressive coding performance and robust cybersecurity features, Mistral’s ML4 is making its mark as a serious contender in the evolving AI marketplace.