Mistral's ML4 Model Breaks Ground in Open AI Coding

Introduction to Mistral's ML4

On October 6, the French AI research lab Mistral unveiled its latest model, the Mistral Large 4 (ML4), affectionately known as "Le Chonk." This new model boasts a staggering trillion parameters and is sparking interest in both coding performance and cyber defense.

Performance Gains in Coding

ML4 has exhibited remarkable improvement, achieving a score of 41% on the Terminal-Bench 4.0, up from 28% in its initial preview. This advancement places it ahead of Opencode's GLM-5.3 by a slender margin. Despite this, ML4 lags slightly behind GLM-5.3 on the overall performance Index, both clocking in at 48, just one point lower than GLM-5.3.

Benchmarks can often be misleading, and it's vital to view these scores with a degree of skepticism. Nonetheless, Mistral's model exemplifies significant enhancements following only a few weeks of additional training, positioning it as a competitor in the economic AI landscape where performance cannot be compromised. Mistral highlights that ML4 uses 47% fewer tokens per task compared to GLM-5.3.

Cybersecurity Capabilities

The launch of ML4 also spotlighted its emphasis on cybersecurity. Mistral has made strides in ensuring the model can be tailored for enhanced data control, a feature increasingly sought after by developers and institutions. Similar to recent initiatives by leaders in the AI industry, Mistral's partnerships and testing focus on bolstering the security features of ML4.

In comparative assessments, ML4 outperformed DeepSeek V4.1 Flash on the CyberGym benchmark, demonstrating its ability to identify and mitigate memory-safety vulnerabilities within C/C++ open-source code. As AI models face scrutiny over security risks, particularly following notable breaches, Mistral is advocating for open models as viable alternatives for robust cyber defense.

Control Over Cyber Defense

Mistral's co-founder, Guillaume Lample, expressed that ML4 is designed with user control in mind, contrasting with proprietary models that may restrict access. This means that enterprises and governments can harness ML4's capabilities without fears of unilateral access revocations by developers.

Aligned with this perspective, Mistral recently joined Nvidia's Open Secure AI Alliance, an innovative collaboration aiming to democratize AI security beyond proprietary frameworks.

Optimizing Resources

Creating a competitive AI model with limited computing resources is a significant challenge for a smaller lab like Mistral. The ML4 model was meticulously trained over two months using 4,000 Nvidia Grace Blackwell GPUs within Mistral's European data centers. For reference, major players like OpenAI reported utilizing about 100,000 GPUs for similar projects.

Conclusion

As Mistral rolls out its new ML4 model, the AI landscape is witnessing a shift where open models like ML4 are beginning to contend seriously with established proprietary designs. With its strong performance in coding and impressive cybersecurity features, ML4 is a noteworthy player in the evolving AI marketplace.