Forge, Mistral AI’s new service for training models on corporate data

General-purpose AI models are trained primarily on public data and designed to handle a wide variety of tasks. Forge takes a different approach: companies operate with internal knowledge — engineering standards, compliance policies, codebases, operational processes — that generic models simply do not have. The platform allows organizations to build models that learn directly from their internal documentation, operational records, and proprietary structured data.

The system covers several stages of the model lifecycle. During pre-training, organizations can develop domain-aware models from large volumes of internal data. Post-training allows teams to refine model behavior for specific tasks and environments. Through reinforcement learning, teams can align models with internal policies, evaluation criteria, and operational objectives.

Forge supports both dense and mixture-of-experts (MoE) model architectures, allowing organizations to balance performance, cost, and operational requirements. It also supports multimodal inputs, combining text, images, and other data formats.

Mistral AI is already working with organizations such as ASML, Ericsson, the European Space Agency, DSO National Laboratories Singapore, and the Home Team Science and Technology Agency (HTX) Singapore, which have trained models on their proprietary data using this platform.

Forge is also designed to be used autonomously by AI agents, without human intervention. An agent can use it to fine-tune models, generate training data, or schedule tasks. The platform also includes evaluation tools to assess model performance before deployment to production.

Target use cases include government agencies needing models adapted to regulatory frameworks or specific languages, financial institutions with compliance requirements, development teams working on proprietary codebases, manufacturing companies with technical specifications and maintenance records, and large enterprises seeking agents capable of operating accurately within their internal systems.

Key points

• Mistral AI launches Forge, a platform for training AI models on proprietary internal data.
• Models learn from each organization's documentation, processes, and records — not public data.
• Covers the full cycle: pre-training, post-training, and reinforcement learning.
• Companies retain full control over their data and models.
• Compatible with dense and MoE architectures; supports text, images, and other formats.
• AI agents can use it autonomously, without human intervention.
• Existing clients include ASML, Ericsson, and the European Space Agency.
• Key sectors: public administration, finance, software development, and industry.