GLM 5.2 hosted by Mistral: digital sovereignty in practice

Since 6 August 2026, Mistral AI's platform has been serving GLM 5.2, the open weights model from the Chinese lab Z.ai. A model designed in Beijing, executed in European data centres, under contract with a French company: the case is almost too perfect an illustration of what digital sovereignty means in practice, once the rhetoric is set aside.

For a lawyer, the question is anything but academic. Every AI tool in the firm, dictation, document summaries, case law research, runs on a model executed somewhere, by someone, under some applicable law. And case file documents passing through an API remain covered by professional secrecy, whether the provider is American, French or Chinese.

Our thesis, against the flag reflex: the nationality of the model is the least important of the three criteria. What matters is who runs the inference, where it runs, and what the contract says. The GLM 5.2 case at Mistral proves it point by point.

Prerequisites before evaluating any AI tool or model

A serious evaluation requires little, but that little is not optional:

• Map what would leave the firm: categories of data (client identities, procedural documents, health data, criminal data), volumes, purposes.
• Gather the prospective provider's contractual documents: terms of use, GDPR data processing agreement (DPA), list of sub-processors. A provider that does not publish them eliminates itself.
• Re-read what your Bar has published on the use of digital tools and AI, before signing anything.
• Block half a day without case files to read those documents. That is the real minimum budget for the exercise.

No technical skills are required. Being able to read a page of API documentation is enough to check a vendor's claims, and a lawyer reads contracts better than anyone.

GLM 5.2 at Mistral: the facts, without the noise

GLM 5.2 has been available on the Mistral platform since 6 August 2026, in public preview, under the identifier zai-glm-5-2. Mistral states that the model is served without modifications: these are the weights published by Z.ai, executed on Mistral's infrastructure, with the same regional controls and service commitments as its own models.

On the model side, GLM 5.2 was released in June 2026 by Z.ai (formerly Zhipu AI), a company based in Beijing, under the MIT licence, with freely downloadable weights. It is a mixture-of-experts model with 744 billion parameters, of which roughly 40 billion are active per token, built for code and long-horizon agentic tasks.

On the platform side, Mistral's announcement of 11 August 2026 goes beyond the model itself: regional endpoints in general availability (inference runs in Europe or the United States, at the customer's choice), a Priority Tier with SLA in public preview, and a coalition of European companies around long-term compute capacity commitments. GLM 5.2 is the first third-party model in this setup, with other open models expected to follow.

The model's seriousness is no longer in question: CAISI, attached to the US NIST, dedicated an assessment to it published in July 2026. When the US administration evaluates a Chinese open weights model, the question of its use arises everywhere, law firms included.

Digital sovereignty plays out across three layers

Digital sovereignty is the ability to control where and by whom your data and systems are processed, and under which law. Applied to AI, this definition breaks down into three distinct layers, which the public debate constantly conflates.

The model layer first. Open weights like those of GLM 5.2 are a file: billions of numbers published under the MIT licence, downloadable and executable by anyone with the right machines. A file sends nothing to anyone. When Mistral runs GLM 5.2, no data passes through Z.ai.

The inference layer next. This is the execution: your questions and documents go in, the answer comes out. Jurisdiction attaches to the operator that executes and to the place of execution, not to the author of the weights. GLM 5.2 called via api.eu.mistral.ai runs in data centres located in the EU and EFTA, operated by Mistral.

The contract layer last. Who is your counterparty, which law applies, which sub-processors are involved, what happens to the content of the requests after processing. This is the layer that settles the matter, and it is the layer a lawyer knows how to read.

flowchart TB
    M["Model layer\nGLM 5.2 weights, MIT licence: a file, nothing more"] --> I["Inference layer\nwho executes and where: Mistral, EU and EFTA data centres"]
    I --> K["Contract layer\ncounterparty, applicable law, sub-processors, retention"]
    K --> V["This is where sovereignty is won or lost"]

Three routes to GLM 5.2, three data sovereignty regimes

The same file of weights produces three distinct legal regimes depending on the access route chosen. A firm wanting to use GLM 5.2 currently has three options: Z.ai's API, Mistral's regional endpoint, or its own servers.

For nearly all Belgian firms, self-hosting a 744 billion parameter model is not a realistic option. The real choice plays out between the APIs, and it illustrates the opening thesis: with an identical model and identical list price, the route determines the counterparty, the place of execution and the contractual framework. Data sovereignty is a property of the route, not of the model.

The concrete trap: believing a European provider is enough

The symptom: a firm picks a European provider, mentally ticks the sovereignty box, and plugs the tool into its case files. We regularly see this shortcut, including among colleagues who are otherwise cautious.

The diagnosis, staying with the Mistral example, whose regional inference documentation is in fact remarkably transparent: the default endpoint api.mistral.ai is global, and Mistral makes no commitment on the inference location for requests sent to it. European localisation requires explicitly calling api.eu.mistral.ai, an option billed at 1.1 times list price. Regional endpoints do not cover everything either: function calling is the only supported tool, and stateful features (Agents, Batch, Files API) are excluded. And the location of inference does not settle data retention, which is a separate setting.

The fix takes half a day: point the configuration explicitly at the regional endpoint, check that every feature you use is covered there, enable zero data retention where eligible, and log the hostname called for each request, as Mistral itself recommends. Without that log, there is no way to demonstrate after the fact where the data went.

None of this is a hidden defect: everything is in the public documentation. The trap is not the provider, it is skim-reading.

Why we would never pick a tool because it calls itself “sovereign”

“Sovereign” is a marketing argument, not a contractual guarantee. The GLM 5.2 case proves it by absurdity: a Chinese model served by a French company under a published DPA can offer more verifiable guarantees on data location than a tool badged 100% European that publishes neither a DPA nor a list of sub-processors. The flag protects nothing; the documents do.

Conversely, geographic suspicion does not hold up any better. Rejecting GLM 5.2 because its weights come from Beijing, while sending your requests to a global endpoint with no localisation commitment, amounts to guarding the wrong door.

Four verifiable questions are enough to evaluate any provider: who is the counterparty? Where does inference run, and is that commitment in the contract? Who are the sub-processors? What happens to the content of the requests after processing? A provider that answers all four with published documents deserves the analysis. A provider that answers with a sales brochure does not.

It is the test we apply to ourselves at Jef: hosting in Europe, GDPR compliance, a published data processing agreement, strictly separated workspaces per lawyer, and independence from model providers (Claude, OpenAI, Gemini, Mistral), precisely so that the choice of model stays separate from the data regime.

What this case changes for a firm's digital sovereignty

The arrival of GLM 5.2 at Mistral makes one shift visible: model weights now circulate independently of the infrastructure that executes them. For a firm, that is good news. The choice of model (quality, price, specialisation) and the choice of data regime (place of execution, contract, sub-processors) become two separate decisions. You can prefer a model for its performance and separately demand a European execution framework; the two no longer conflict.

The flip side: the geographic argument from authority is no longer enough, in either direction. A firm's digital sovereignty is built contract by contract, endpoint by endpoint, request log in hand. It is lawyer's work more than IT work, and that is rather reassuring.

Jef (https://www.jef.chat/en) applies this logic to the daily work of Belgian lawyers: an AI assistant built on the firm's own document base, with sourced answers, strictly separated workspaces, hosting in Europe and full independence from model providers. Signing up is free (50 messages included).