Mistral, one of Europe’s leading AI developers based in Paris, has unveiled Mistral Large 4 (ML4), an open-weight model the company plans to release by the end of the month. Mistral highlights that ML4 was specifically tuned for cybersecurity tasks and produces fewer refusals than some proprietary systems such as Anthropic’s and OpenAI’s models.
“ML4 pairs top-tier cyber performance with open weights and self-deployment, giving organizations both the capability and the autonomy to run advanced security work under their own policies,” the company states in its announcement.
Training and infrastructure
ML4 was trained in Mistral’s European datacenters on 3,800 NVIDIA Grace Blackwell GPUs. According to the company, this scale allowed them to refine the model’s cyber-oriented capabilities while offering open weights and self-deployment options for organizations.
Performance and benchmark scores
The published comparisons suggest that upon release ML4 will be the strongest Western open-weight model available, although some Chinese models remain more powerful on several benchmarks. Selected scores cited by Mistral include:
- Artificial Analysis – DeepSWE 1.1: ML4 = 62, Kimi K3 = 68, GLM-5.3 = 61.
- Terminal-Bench 4.0: ML4 = 28.3, GLM-5.3 = 41.9.
- Finance Agent V2: ML4 = 54.7, GLM-5.3 = 55.8.
An internal preference test (vibes eval) reported that users preferred ML4 for CAD and STEM tasks, while ML4 performed on par or close to GLM-5.3 in finance and coding tasks.
Why this matters
The ML4 release is significant for European technological sovereignty: Mistral is seen as the continent’s best candidate to serve as a multi-national AI champion. The announced results, however, underscore that competing with leading Chinese models will remain challenging.
Mistral plans to make the ML4 open-weight release available soon, enabling organizations to deploy and operate the model under their own policies and security requirements.
