Mistral Launches ML4: A Powerful Open Model for Coding
Mistral has unveiled its ML4 model, also known as Le Chonk, showcasing impressive coding performance and enhanced cybersecurity features.
On October 6, the French artificial intelligence lab Mistral released its latest machine learning model, Mistral Large 4, affectionately named "Le Chonk." This groundbreaking model, boasting a staggering trillion parameters, signifies a remarkable advancement in the race of open-source AI models.
Coding Performance Breakthroughs
The coding capabilities of ML4 have made headlines, particularly due to its recent performance on the Terminal-Bench 4.0 coding benchmark. The model's score increased significantly to 41%, climbing from 28% during its preview phase. This leap in performance enables ML4 to outpace Opencode’s GLM-5.3 by a narrow margin, reflecting its growing competitiveness in the AI model arena.
In the context of overall performance, ML4 scored 48, which places it just behind GLM-5.3 and Codex GPT-6.1 Terra, both scoring 49. Nevertheless, it's important to interpret these benchmark scores cautiously, as they can vary based on the methodologies utilized by different models.
Strengthening Cybersecurity Measures
Mistral has prioritized cybersecurity within the ML4 model, offering customizable controls to enhance user security. As AI applications expand, ensuring robust security measures has become increasingly crucial.
Initial assessments of ML4's cybersecurity capabilities indicate its effectiveness in the CyberGym benchmark, designed to evaluate models on their ability to identify and rectify memory-safety vulnerabilities in open-source C and C++ code. ML4 has reportedly outperformed DeepSeek V4.1 Flash, showcasing strong qualifications in cybersecurity.
The security challenges faced by proprietary models have prompted increased interest in open models like ML4, which are thought to offer increased user control amidst rising concerns. Mistral co-founder Guillaume Lample emphasized that ML4 is engineered for security under user management, contrasting the risks associated with proprietary models that can revoke access at any time.
Efficient Training in a Competitive Landscape
Creating a leading AI model like ML4 in a resource-constrained environment is a notable achievement for a company like Mistral, which lacks the vast resources of larger organizations like OpenAI. ML4 was trained over two months using 4,000 Nvidia Grace Blackwell GPUs located in Mistral’s European data centers.
For perspective, OpenAI’s recent advancements reportedly required around 100,000 GPUs for training their models. This efficiency underscores Mistral's capability to maximize resources while developing high-quality AI tools.
Conclusion: A New Era of Open AI Models
The launch of ML4 positions Mistral prominently in the ongoing competition between open-source and proprietary AI technologies. With its impressive coding capabilities and dedicated focus on cybersecurity, Le Chonk stands as a formidable option in the evolving landscape of artificial intelligence.
