Mistral Unveils Le Chonk: A Game-Changer in Open AI Models
Mistral's new ML4 model, known as 'Le Chonk', showcases impressive coding skills and cybersecurity features, enhancing open AI options.
The French AI lab Mistral has made a significant stride in the field of machine learning by unveiling its latest model, Mistral Large 4 (ML4), also affectionately named 'Le Chonk'. This model boasts a phenomenal one trillion parameters and was released on October 6. Mistral's launch highlights the growing competitiveness of open models in the AI landscape, particularly as they rival their proprietary counterparts.
Impressive Benchmarking Results
One of the noteworthy aspects of ML4 is its coding performance, which has been quantified through various benchmarks. In a recent assessment using the Terminal-Bench 4.0 coding benchmark, ML4's score surged from 28% to 41%, surpassing Opencode's GLM-5.3 by a single point. However, in the overall Index, which combines scores across three benchmarks, ML4 achieved a score of 48, placing it just behind GLM-5.3 and Codex GPT-6.1 Terra, both scoring 49.
It's essential to approach these benchmark figures with caution, as the evaluations can vary and some models may use unfair advantages to boost their scores. Nevertheless, Mistral's improvements showcase significant advancements in just a few weeks of additional training, indicating that open models like ML4 are becoming more viable options for users looking for cost-effective alternatives without compromising on performance.
Focus on Cybersecurity
Alongside its coding capabilities, ML4 emphasizes robust cybersecurity features, designed to bolster user control and data protection. In light of recent security concerns, Mistral has collaborated closely with developers to enhance and assess the model's capabilities in dealing with potential threats.
The initial evaluations show promise for ML4 in the realm of cybersecurity. Using the CyberGym benchmark, which measures the ability to identify memory-safety vulnerabilities in C/C++ code, ML4 has outperformed DeepSeek V4.1 Flash. This performance signifies a notable achievement, especially as open models like ML4 are increasingly seen as a more secure option amidst rising risks associated with proprietary models.
Guillaume Lample, Mistral's co-founder, noted that ML4 is built with user security tightly integrated, allowing enterprises and governments to effectively safeguard against attacks that leverage vulnerabilities in closed models.
Resourceful Development Amid Constraints
Mistral's ability to build an influential AI model without the considerable resources available to larger entities like OpenAI underscores the efficiency of its approach. ML4 was trained from the ground up over two months on a fleet of 4,000 Nvidia Grace Blackwell GPUs located in their European data centers.
For context, OpenAI trained models like GPT-6 Astra reportedly using around 100,000 GPUs. This contrast highlights Mistral's innovative capacity to maximize its resources while delivering high-performance AI tools.
A Competitive Future
The debut of ML4 positions Mistral as a key player in the battle between open-source and proprietary AI models. With its strong coding abilities and an emphasis on security, Le Chonk is set to make waves in the evolving AI environment.
