Mistral to host external models, GLM first
The French AI company Mistral is taking a step that broadens its position from being purely a model maker to becoming a supplier of infrastructure.
Users will be able to use, alongside Mistral’s own models, those of third parties. These are open models and the first to be installed are Z.ai’s GLM models. They come from China, which is making a principled choice for openness in order to outmanoeuvre American giants.
Customers will soon be able, for example, to use Mistral models for general tasks and GLM-5.2 for complex coding or agent workflows, without needing a separate cloud provider, API integration and compliance procedure for each model.
In this way, a European company uses a Chinese open model, while processing runs through a European Mistral region.
This is not unpleasant for users because they are not locked into Mistral’s model ecosystem. They can also start choosing others. These will presumably also be Chinese models, such as Kimi, DeepSeek, Qwen and MiniMax.
Mistral is not the only one opting for such a multi-model approach, incidentally. Amazon has been doing this for some time with Bedrock. Bedrock offers a single API, a single billing and security layer and access to models from various providers, such as Anthropic, Meta, Cohere, Mistral and Amazon itself. AWS supplies the infrastructure on which the models are offered.
Microsoft is also doing something similar with Azure AI Foundry. Models from Microsoft itself, Anthropic and other partners, including open models, are available there.
Mistral also announces that it is forming a coalition that pools demand for AI computing power. The purchased capacity is recorded in European Compute Units, or ECUs. These give participants access for several years to infrastructure that Mistral is building. The capacity can be used for, for example, inference, training and model adaptations.
Names mentioned as participants include Amadeus, ASML, Capgemini, CMA CGM, Factory and Groupe Caisse des Dépôts.
The concrete infrastructure objective is 200 megawatts of European capacity by the end of 2027 and ultimately up to 1 gigawatt by the end of 2030.
This development is a second signal that Mistral is moving from model maker to supplier of AI infrastructure.
Emerce
Subscribers 0
Fans 0
Followers 0
Followers