AI CRITIQUE

Görsel, büyük ve endüstriyel bir odada uzun bir sıra halinde dizilmiş koyu renkli sunucu dolaplarını, gri ve mavi tonlarında stilize bir çubuk grafik ile geometrik şekillerin üzerine bindirilmiş halini göstermektedir.

Mistral Large 4 tested as the best open model outside China

Read the article English (automatic translation)

Summary

Mistral Large 4, an open-weights model with 1.05 trillion parameters trained on 3,800 Grace Blackwell GPUs in European data centers, has been released as a preview. Ranked 8th among open models in independent benchmarks, it trails six Chinese competitors but leads all Western open-weights models by a significant margin.

Why it matters

This is the most concrete test of whether European AI sovereignty can produce frontier-level open models without relying on Chinese weights. An 8-point gap compared to Xiaomi's MiMo-V2.6-Pro highlights the remaining distance to cover. Since reinforcement learning training is still ongoing and the license name is unspecified, the preview data serves as a lower bound. This indicates that the final open-weights release will be a genuine competitive event for the non-Chinese ecosystem.

Highlights

Ranked 8th on the Artificial Analysis intelligence index (38 points): 8 points behind the top-ranked Chinese open model, 12 points ahead of the best US open model, and 4 points ahead of South Korea's Motif 3 model.
Trained from scratch on 3,800 Grace Blackwell GPUs in approximately 2 months. Reinforcement learning training is still ongoing. The training is supported by the €3 billion in funding raised in September.
Cybersecurity: Ranked 5th out of 18 models on the Cyber Index. However, Mistral's own platform still lists the Chinese-origin GLM 5.3 as a hosted third-party model.

Critical take

Mistral's claim of being the 'most capable open model outside China' is technically defensible but strategically convenient. The model is only 4 points ahead of a Korean competitor. The license remains unnamed, listed simply as 'Open'. Moreover, the company hosts Chinese-origin GLM weights on its own servers, which undermines the sovereignty narrative it sells. Since reinforcement learning training is incomplete and weights have not yet been published, the preview benchmarks are provisional. Characterizing a single data point as a generational leap carries the risk of exaggeration.

Bu kritik, yerel yapay zeka modeli Qwen3.8-27B tarafından yazılmıştır. Haberi ben seçtim; kritik metnini yayımlanmadan önce tek tek denetlemiyorum.This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.

Back to the issue