Mistral Large 4 trails Chinese open models in tests
The AI landscape is heating up, and Mistral has just thrown down a gauntlet, launching its trillion-parameter Large 4 (ML4) model, claiming it is pushing the frontier of open-weight performance. This isn’t just another model release; it’s a bold statement about what open-source AI can achieve, and it immediately entered the global benchmarking arena.
Independent analysis from Artificial Analysis (AA) placed ML4 at a respectable 38 on the Artificial Analysis Intelligence Index, positioning it as the most intelligent model available from outside the US and China. This achievement puts Mistral firmly in the spotlight, challenging the dominance of closed models and established open-source contenders.
But the real story lies in the specifics of the performance. When looking at cybersecurity and complex reasoning, ML4 demonstrated exceptional capability. It scored 82% on the demanding CyberGym-E2E-AA measure, leading competitors like OpenAI GPT-6 Luna (max, 78%) in this specific test. This suggests that ML4 excels not only in general intelligence but also in critical, real-world security applications.
The model’s multimodal abilities are also a major draw. Mistral Large 4 Preview handles both text and image input, boasting a context window of 512k tokens. Furthermore, improvements in its API allow it to process up to 100 images per request, significantly boosting its ability in document and image reasoning, showing an impressive 18-point gain over its predecessor.
Crucially, Mistral is not just focusing on intelligence; they are focusing on efficiency. The cost of running ML4 is remarkably competitive. While the list price stands at $1.13 per Intelligence Index task, a 50% launch discount brings the cost down to $0.57. This affordability allows ML4 to offer high-end performance at a fraction of the cost of some of its larger counterparts.
In the broader competition, the leaderboard is tight. Other leading open models, including those from Chinese developers like DeepSeek V4.1 Flash and Kimi K3, also score highly on the intelligence index, reflecting the fierce competition in the open-weight space. However, Mistral’s performance positions it strongly against these rivals, demonstrating resolve and the potential for niche application.
The foundation for this success is rooted in massive infrastructure. ML4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers, underscoring the company’s commitment to scaling specialized compute capacity. As Mistral plans to release the model’s open weights by the end of October, the focus now shifts to seeing if this performance translates into widespread adoption and whether it can truly close the performance gap with the largest closed models.
Credit
