Mistral Large 4 trails four Chinese open-weight models on a composite index

Four Chinese open-weight variants score higher on the same index: Xiaomi’s MiMo-V2.6-Pro at 46, Z.ai’s GLM-5.3 (max) at 45, Moonshot’s Kimi K3 (Max) at 44, and DeepSeek V4.1 Flash (Max) at 39. Those results apply to the specified variants, not every version or configuration from each developer.

Artificial Analysis also estimates a cost of $1.13 per Intelligence Index task for Large 4, compared with $0.13 for MiMo-V2.6-Pro and $0.27 for DeepSeek V4.1 Flash. Those are evaluator-calculated estimates based on listed API prices, not a guarantee of what an individual user will pay.

Where do the other tests point?

Frandroid reported that Large 4 placed sixth among 76 models on Vals.ai’s Harvey Legal Agent Benchmark, scoring 15.83%. It ranked first among open-weight models in that evaluation. The result indicates how the model performed on that benchmark; it does not establish that it will handle legal work reliably in practice.

On Artificial Analysis’s CyberGym-E2E test, Large 4 scored 82%, Frandroid reported. Its Cyber Index score was 50, tying GLM-5.3-Flash and trailing MiMo-V2.6-Pro’s 56. Mistral told Frandroid that some closed models score near zero on CyberGym-E2E because they refuse the task, a caveat that makes the test harder to read as a broad measure of cybersecurity ability.

The results were less favorable on agent and terminal-coding evaluations. Frandroid reported that Large 4 ranked 43rd of 52 models in Agent Arena, with a net score of −6.6%, using data dated Oct. 8. On Terminal-Bench 4.0, reported scores differed: Artificial Analysis recorded 26.8%, Vals.ai 22.73%, and Mistral 28.3%. The spread is a reminder that results can depend on the evaluator and test setup.

Mistral says Large 4 is competitive with leading open models, particularly in cybersecurity, finance and law. At the launch, the company’s chief scientist, Guillaume Lample, said the model’s cyber-defense capabilities could help enterprises and governments defend against attacks. That is the company’s rationale; the benchmark results do not independently establish effectiveness in real-world security work. Euronews reported on the launch.

Can users download or host Large 4?

Not yet. Mistral announced a public preview through its API on Oct. 6 and said the weights would follow by the end of October. As of Friday, Artificial Analysis lists the preview as proprietary and the weights as not public.

That distinction matters for users choosing an open-weight model. API access lets users try the preview through Mistral’s service; without published weights, they cannot evaluate the released model for self-hosting. The license and deployment requirements will also matter once the weights arrive. Until then, the benchmark comparisons offer evidence about specific tasks, while leaving local use and performance on an individual team’s workload unresolved.