Mistral, from sovereign laboratory to host of Chinese models

A column in the Journal du Net retraces the shift of Mistral: from the training of original models to the promotion, on its own platform, of the Chinese GLM 5.2 model.

In a column published on August 17, 2026 on the Journal du Net, Pierre-Louis Biojout (NanoCorp) questions the trajectory of Mistral AI. Three years after its acclaimed debut — Mistral 7B and Mixtral, inventive architectures that forced the industry to look to Paris — the author believes that the company has changed its business without saying so.

Architecture designed in Hangzhou

First signal noted: Large 3, the flagship model released in December 2025, would use the configuration of DeepSeek V3 workstation for workstation — 7,168 of hidden dimension, 61 layers, 3 dense layers before the MoE layers, and the same latent multi-head attention to the compression rows. An important nuance that the chronicle is keen to ask: the weights remain those of Mistral, trained on its own data, in its own data centers, with an in-house tokenizer of 131,000 tokens (compared to 129,000 at DeepSeek). So it's not a rebadged Chinese model, but the architecture plan comes from elsewhere.

GLM 5.2 Highlighted Without Changes

The next step dates from August 11, 2026: GLM 5.2, developed by the Chinese laboratory Z.ai, appears as a model highlighted on the Mistral platform. The model is offered in its original version, including the name of the Chinese laboratory: no retraining, no adaptation, Mistral only intervenes at the stage of model execution. Technical director Timothée Lacroix justified it to VentureBeat by the quality of the model and its open weight nature: there was, in his opinion, no reason to deprive oneself of it. The author sees it as a response from a hosting provider, not a research laboratory.

A commercially defensible pivot

The column acknowledges the economic logic of the shift: multi-year contracts for computing capacity, a target of one gigawatt in Europe by 2030, Microsoft leasing capacity in Mistral's data centers, and a revenue target of one billion euros over the year. Job offers accompany the movement, with pre-sales profiles, solutions engineers and forward deployed engineers rather than researchers. The conclusion of the column focuses on the shift in the meaning given to the word sovereignty: building the models yourself and hosting them on behalf of other laboratories are not the same business, even when the servers are European in both cases.