Does Mistral Large 4's cybersecurity prowess outweigh lagging intelligence scores?
Independent tests show Mistral Large 4 leads non-Chinese models but trails several open-weight Chinese peers in overall intelligence scores.
Key points
- Mistral Large 4 scored 38 on the Artificial Analysis Intelligence Index, ranking highest among models from outside the US and China.
- Five Chinese open models, including MiMo-V2.6-Pro at 46 and GLM-5.3 at 45, achieved higher overall scores than Mistral's new release.
- Mistral Large 4 achieved an 81.7% score on the CyberGym-E2E-AA cybersecurity measure, outperforming competitors in that specific metric.
Mistral launched its trillion-parameter Large 4 model this week, claiming it pushes the frontier of open-weight performance. According to Tom's Hardware, independent benchmarking firm Artificial Analysis ranked the model as the best available from outside the United States and China. However, several Chinese open models scored higher on the overall index, highlighting a competitive gap in general intelligence metrics.
How does Large 4 compare to Chinese models?
Large 4 scored 38 on the Artificial Analysis Intelligence Index v4.3.2. At least five Chinese open models surpassed this score. DeepSeek V4.1 Flash (max) scored 39, while GLM-5.3-Flash reached 42. Kimi K3 (max) achieved 44, GLM-5.3 (max) hit 45, and MiMo-V2.6-Pro scored 46. These results place Large 4 behind these specific Chinese offerings in the broader intelligence ranking.
View the original post on X
What are the size and cost differences?
Mistral Large 4 is a mixture-of-experts model trained on Nvidia’s Grace Blackwell GPUs. It has 49 billion to 52 billion active parameters. In contrast, GLM-5.3-Flash has only 320 billion total parameters and 18 billion active. Cost-wise, Large 4 lists at $1.13 per task, or $0.57 with a launch discount. MiMo-V2.6-Pro costs $0.13 per task, and GLM-5.3-Flash costs $0.25, making the Chinese options significantly cheaper.
How does it perform in cybersecurity tests?
Large 4 scored 81.7% on the CyberGym-E2E-AA end-to-end cybersecurity measure. Artificial Analysis called this its strongest result, and Mistral stated it is the highest of any model. This score beats GLM-5.3-Flash, which scored 74%. However, Large 4 scored only 16% on DeepsecBench-AA and 51% on CWE-Bench-AA. Artificial Analysis projects it will rank among the top three open-weight models on the Cyber Index once weights are released.
When will the model weights be available?
Mistral stated that the model weights are due by the end of the month. Until then, the model is not fully open. Mistral bets that Large 4 is state-of-the-art among open models for enterprise workloads in cybersecurity, finance, and law. The company also claims its 1.6 billion-parameter vision encoder can surpass frontier closed models in visual grounding tasks.
What to do and how to stay safe: Mistral
- Monitor official announcements from Mistral for the release of model weights to verify open-weight status.
- Compare cybersecurity benchmark scores like CyberGym-E2E-AA against your specific security testing requirements.
- Evaluate cost-per-task metrics relative to the performance gains offered by cheaper alternative models.
- Test visual grounding capabilities independently if your workflow relies on high-precision image-text alignment.
Step-by-step guide: How AI Risk Management Works: The Hidden Failure Points
General security guidance from the Sandbox Dispatch newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which model scored highest on the CyberGym-E2E-AA benchmark?
Mistral Large 4 scored 81.7%, which Artificial Analysis called its strongest result and the highest of any model.
How much does Mistral Large 4 cost per task?
It costs $1.13 at list price or $0.57 with the 50% discounted launch price, according to Tom's Hardware.
Are the model weights available now?
No, Mistral says the weights are due by the end of the month, so the model is not yet fully open.
