Mistral Unveils Forge at GTC 2026, Letting Enterprises Train Custom AI Models from Scratch
The French AI startup's new platform challenges the fine-tuning approach favored by rivals by enabling governments and corporations to build proprietary models on their own data with embedded engineering support.
Mistral AI used the stage of Nvidia's GTC 2026 conference in San Jose on Tuesday to announce Mistral Forge, a platform that allows enterprises and governments to build custom AI models trained from scratch on their own internal data rather than relying on fine-tuning or retrieval-augmented generation approaches. The announcement positions Mistral as a bespoke model builder targeting use cases where off-the-shelf AI systems fall short due to language, compliance, or domain-specificity constraints.
Forge customers can select from Mistral's library of open-weight models — including the recently introduced Mistral Small 4 — as starting points, and then build out fully customized versions using proprietary documents, workflows, and institutional knowledge. Mistral's approach is differentiated by the deployment of what it calls forward-deployed engineers, who embed with customers to surface appropriate training data, design evaluation frameworks, and adapt the platform to the customer's specific technical and operational needs.
This high-touch model reflects Mistral's belief that custom AI development at the enterprise level requires domain expertise that cannot be fully automated.
The platform is already in use at a notable set of early adopters: Swedish telecommunications giant Ericsson, the European Space Agency, Italian consulting company Reply, and two Singaporean government defense and technology agencies — DSO and HTX. ASML, the Dutch lithography equipment maker and a lead investor in Mistral's Series C round last September at a valuation of approximately 11.7 billion euros, has also adopted Forge.
Mistral head of product Elisa Salamanca told TechCrunch that Forge is designed for four types of customers: governments that need models tuned to their national language and cultural context; financial institutions with strict regulatory compliance requirements; manufacturers seeking deep customization for industrial processes; and technology companies that want models optimized for their proprietary codebases. The announcement reinforces Mistral's identity as a European AI lab committed to offering an alternative to U.S.-controlled foundation model infrastructure, particularly for institutions with sovereignty and data residency concerns.
Read the original reporting at TechCrunch.
