Mistral AI now supports third-party open models on its platform, beginning with Z.ai's GLM-5.2. The move broadens Mistral's offerings beyond its own model family, running third-party models on its existing infrastructure.
The company is advancing an AI sovereignty strategy, aiming to give enterprises and governments control over models, infrastructure and compute capacity. The approach targets regional compliance and reliability requirements for mission-critical workloads.
Mistral introduced Regional Endpoints, now generally available, allowing customers to choose inference locations in Europe or the U.S. The feature helps align processing with data-residency, regulatory and latency requirements. Limited, safeguarded data transfers to sub-processors outside the selected region may still occur.
A new Mistral Priority Tier is in public preview, offering committed service levels for critical operations. The tier includes custom rate limits and is backed by an uptime service level agreement. Mistral said it is the only European AI lab offering both regional processing choice and an SLA-backed service.
Beyond inference controls, Mistral plans to build up to one gigawatt of compute capacity in Europe by 2030. The infrastructure is intended to provide scalable, sovereign AI capabilities that keep data and model control with customers.
Many customers already run Mistral models inside their own data centers and cloud environments. As AI moves deeper into production, those customers require confidence in capacity availability, resilience and regional controls. Mistral said it will complement customer-managed infrastructure with capacity it directly operates.
The company emphasizes open model weights as a core differentiator. Open weights let customers inspect models, adapt them and retain the intelligence built within those systems. Mistral participates in the Open Secure AI Alliance and the NVIDIA Nemotron Coalition.
Z.ai, formerly known as Zhipu, is a Chinese AI lab. Its GLM-5.2 is a third-party open-source text model hosted by Mistral without modifications, built for long-context coding and agentic workflows.
GLM-5.2 carries a one million token context window, placing it among the longer-context models available on the market. Its Mixture of Experts architecture, with 40 billion active parameters per token, is designed to manage inference costs efficiently.
While GLM-5.2 is open source, access on Mistral's hosted platform is not free. A GLM Coding Plan is required, with a Lite tier starting at approximately $18 per month.
Hosting a model from a Chinese AI lab sits in tension with Mistral's European sovereignty positioning—a tradeoff the company will have to defend as it courts government and enterprise customers for whom provenance is a compliance question, not a marketing one.
