Mistral's new trillion-parameter Large 4 model scored 38 on Artificial Analysis' Intelligence Index v4.3.2, placing it behind at least five Chinese open-weight models including Xiaomi's MiMo-V2.6-Pro at 46. The independent benchmarking firm nonetheless ranks it as the best model available from outside the U.S. and China.

Mistral launched the mixture-of-experts model this week, claiming it "pushes the frontier of open-weight performance." The model has 1.05 trillion total parameters with 49B–52B active, and Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters — up from the 3,000 H200s used for last year's 675B-parameter Large 3. Training took roughly two months, as reported by VentureBeat.

Where does Large 4 actually win?

Its strongest result is cybersecurity. ML4 scored 81.7% on AA's CyberGym-E2E-AA end-to-end test, which AA calls its "strongest result" and Mistral calls "the highest of any model." That figure beats OpenAI's closed GPT-6 Luna (max) at 78%, and it came from a post-training run Mistral says is "still in flight."

The picture narrows on the other two tests that make up AA's Cyber Index. ML4 manages only 16% on DeepsecBench-AA and 51% on CWE-Bench-AA, for an overall Cyber Index score of 50 — level with Z.ai's GLM-5.3-Flash and below Grok 4.7 (xhigh) and MiMo-V2.6-Pro, both at 56.

CEO Arthur Mensch went further at an event in Abu Dhabi, saying the model is "above the Chinese models on certain aspects, including cyber," Reuters reported. Mistral adds that ML4 solves 93% of Cybench, a set of 40 challenges drawn from security competitions.

What does the cost gap look like?

Running AA's index on ML4 costs $1.13 per task at list price, or $0.57 at the discounted launch price. That compares with:

• $0.13 for Xiaomi MiMo-V2.6-Pro, comparable in total size to ML4
• $0.25 for GLM-5.3-Flash, which is about one-third the size at 320B total parameters
• $0.07 for OpenAI's GPT-6 Luna (max), which matches ML4's index score of 38

The cost stems from volume: ML4 needed 200 million output tokens to run AA's index, against a median of 81 million. AA's index-versus-cost chart places ML4 below the Pareto line and just outside the "most attractive quadrant."

Speed is a partial offset. ML4 generated 116.1 tokens per second with 1.46 seconds to first token, roughly 33% faster than its tier median and 2.6x faster to first token, measured on Mistral's own API.

Are the weights actually open?

Not yet. Unlike every Chinese rival above it, ML4 remains proprietary, with weights due "by the end of the month." AA projects that once the weights ship, ML4 "will rank among the top three open weights models on the Cyber Index." Weights, architecture details and additional benchmarks are expected before the end of October, at which point AA may re-score the model.

Mistral positions ML4 as "state-of-the-art among open models" on enterprise workloads such as "cybersecurity, finance and law," and claims its 1.6B-parameter vision encoder lets it surpass "even frontier closed models" in visual grounding. Mistral's own figures include DeepSWE v1.1 at 61.7%, Coding Agent Index at 49.8%, AutomationBench at 59.9% and Dense 200 at 42%.

The company raised a €3 billion Series D, which it calls "the largest equity round ever raised by a European technology company," and Mistral Compute has planned for 18,000 Grace Blackwell chips under its announced Nvidia partnership. ML4 will serve as the base for "a new generation of specialized and optimized Mistral models," with open weights runnable on private cloud or on-premises once released.

The commercial test comes when the weights arrive: an eight-point index gap against a similar-size Chinese model that costs nearly nine times less per task defines the competitive problem Mistral must close with its specialized follow-on models.