Mistral launches Forge to let enterprises train AI models from scratch on their own data

Mistral launched Forge, an enterprise platform enabling companies to train custom AI models on proprietary data rather than layering data via RAG or fine-tuning existing models. Announced at Nvidia GTC, Forge is already live with Ericsson, the European Space Agency, and ASML — the chip equipment maker that led Mistral's Series C at an EUR 11.7B valuation.

Most enterprise AI deployments hit a ceiling because general-purpose models cannot internalize domain-specific workflows and terminology. Forge removes that ceiling by fundamentally retraining the model rather than bolting retrieval onto it — a meaningful architectural difference for companies with complex, non-English, or operationally sensitive data. Mistral is on track to exceed $1B ARR and is positioning Forge as the enterprise alternative to OpenAI and Anthropic: more data control, no vendor lock-in on model versions, and the ability to keep proprietary intelligence in-house.