Mistral offers the GLM model from the Chinese lab Zhipu AI on its platform, with open access (open weights). This choice, which sparked reactions in the French tech scene, is, according to the source, a bet that is both financial and technical, and it paved the way for the arrival of Mistral Large 4, a 1-trillion-parameter model released in early October.
- Mistral propose le modèle chinois GLM de Zhipu AI, en accès libre, ce qui a suscité des débats.
- Le choix est surtout financier : il génère des revenus récurrents en facturant chaque réponse produite.
- Il vise aussi l’IA souveraine, car les entreprises regardent où l’inférence s’exécute plutôt que l’origine du modèle.
- Il permet à Mistral d’apprendre à faire tourner des modèles de pointe, Mistral Large 4 restant le modèle par défaut.
Some saw it as a betrayal of the company’s sovereignty promise, others as a shift towards a consulting and cloud business. Essentially, the explanation is primarily financial. AI developers spend enormous sums on physical infrastructure and graphics processing units (GPUs) to train their models. In Anthropic’s IPO filing, cited by the source, $7.33 billion was spent on compute in 2025, against $4.6 billion in revenue.
Mistral has far fewer resources than OpenAI or Anthropic, despite several successive funding rounds, including €1.7 billion in 2025. Integrating GLM thus aims to create a recurring revenue stream by charging companies for each response generated within their services. “It’s clearly a very good business, bringing recurring revenue with amortized capex, ” explains Bilal El Alamy, co-founder and CEO of Pyratz (venture builder).
An Offer Designed for Sovereign Inference
Mistral is also inspired by Palantir and its engineers embedded with clients, known as ‘forward deployed engineers.’ On-site, these engineers install and deploy technologies and advise companies on AI in general. This presence allows Mistral to establish itself with large accounts, particularly around sovereign AI, a segment where Marc Trilling, president and co-founder of Saegus, sees “strong growth prospects.”
According to Saegus, which advises CAC 40 companies, these firms are less interested in a model’s origin than in where its inference runs, meaning the process that makes it operate on servers. Mistral states in an August 11 blog post that GLM and future third-party models “will run on the same infrastructure, with the same regional controls and the same service commitments” as its own models.
This demand meets a concrete need. Nicolas Gaudemet, chief AI officer at Onepoint, observes that “some of our clients have real sovereignty issues, ” especially in the public sector with highly sensitive data. For many companies, running these models on-premise or on their own cloud infrastructure remains too costly, as current models are very heavy. In his view, Mistral’s interface then becomes the most logical solution. He cites Scaleway, OVHcloud, Cloud Temple, and Mistral as the only players currently offering sovereign API access, at scale, to frontier models.
A Test Bed for Frontier Models
Hosting GLM also gives Mistral the opportunity to learn how to run a frontier model at scale on its own infrastructure. Properly serving thousands of clients with models of several hundred billion parameters is tricky, and OpenAI and Anthropic have faced challenges, even though their customer base is incomparably larger. Bilal El Alamy believes that once the infrastructure is in place, Mistral can catch up with its competitors by leveraging existing best practices.
Mistral Large 4 still lags behind the most performant Chinese and American models. However, training a 1-trillion-parameter model with only 4,000 Nvidia Grace Blackwell chips is, according to the source, a technical success. “We really think we’ve found the right method to iterate, ” summarizes Pierre Stock, VP of Science at Mistral AI.
The compute capacity purchased with the Series C funding round, announced in September 2025, will only be fully operational between late October 2026 and early 2027. Until then, GLM is not meant to remain in the foreground. “The plan is obviously to have ML4 as the default model in all our products, ” assures Pierre Stock. Guillaume Lample, co-founder and chief science officer of Mistral, clarifies that the Chinese model will remain in the catalog, but that “by default, the goal is obviously to promote Mistral Large 4.”
