On October 6, Mistral, a French AI lab, unveiled its latest machine learning model, Mistral Large 4, known affectionately as "Le Chonk." This remarkable model boasts a trillion parameters and marks a significant achievement in the competition among open-source AI models.
Achieving New Heights in Coding Performance
The coding capabilities of ML4 have seen a notable improvement. In the recent Terminal-Bench 4.0 coding evaluation, it soared to a score of 41%, up from 28% in its prior preview version. This advancement has allowed ML4 to edge out Opencode's GLM-5.3 by a marginal point, showcasing its growing prowess in the competitive AI landscape.
However, while ML4 has demonstrated these capabilities, it remains slightly behind both GLM-5.3 and Codex GPT-6.1 Terra in the overall performance index, with scores of 48 compared to their 49. Benchmark scores can be subjective, and it's crucial to evaluate them critically, especially considering that some models might employ dubious methods to enhance their results.
Cybersecurity: A Top Priority
Mistral has placed significant emphasis on the cybersecurity features of ML4. The company has designed the model to provide customizable control for users, addressing the increasing need for strong security measures in AI systems.
Initial tests of ML4's cybersecurity performance reveal that it excels in the CyberGym benchmark, which evaluates a model's proficiency in identifying and resolving memory-safety vulnerabilities in open-source C and C++ projects. ML4 outperformed DeepSeek V4.1 Flash in its tests, highlighting its robust capabilities in the realm of cybersecurity.
With security concerns rising, open models like ML4 are being viewed as viable alternatives to proprietary systems, which have faced various security challenges in recent years. Mistral's co-founder, Guillaume Lample, stated that ML4 was engineered to maintain user control, countering the risks present in proprietary models that may restrict access at any moment.
Efficient Resource Utilization in Model Training
Developing a cutting-edge AI model while navigating a resource shortage is a noteworthy accomplishment, particularly for a lab like Mistral, which operates on a smaller scale compared to larger entities like OpenAI. The training of ML4 was conducted over a two-month period utilizing 4,000 Nvidia Grace Blackwell GPUs in Mistral’s European data centers.
In contrast, larger companies, including OpenAI, reportedly require upwards of 100,000 GPUs for training their advanced models, making Mistral's achievement in efficiency even more impressive.
Conclusion
With the launch of ML4, Mistral is carving out a significant position in the ongoing rivalry between open-source and proprietary AI. With its enhanced coding capabilities and strong security features, Le Chonk stands as a compelling option in the rapidly evolving AI marketplace.
