Introducing Forge

Forge is a system that enables companies to build frontier-level AI models using internal knowledge rather than public data. It puts information accumulated within an organization—such as engineering standards, compliance policies, codebases, and operational procedures—at the center of model training.

Forge is already collaborating with organizations such as ASML, DSO National Laboratories Singapore, Ericsson, European Space Agency, HTX Singapore, and Reply, training models tailored to their complex systems and core technologies using each organization's proprietary data.

Forge supports the entire model lifecycle.

• Pre-training: Building domain-aware models with large-scale internal data
• Post-training: Refining behavior for specific tasks and environments
• Reinforcement learning: Aligning with internal policies, evaluation criteria, and operational goals to improve agent performance

This approach goes beyond simple answer generation, enabling enterprise agents that understand internal terminology and procedures, select tools more accurately, and reliably perform multi-step workflows. As a result, domain-specific models function as operational components that reflect an organization's policies and business logic.

On the architecture side, Forge supports both dense and mixture-of-experts (MoE), and handles multimodal input when needed. Dense focuses on broad general-purpose performance, while MoE focuses on operating larger models with lower latency and compute cost.

Forge is designed to be agent-first. Autonomous agents like Mistral Vibe can perform hyperparameter search, job scheduling, synthetic data generation, and even eval hill-climbing, while Forge monitors metrics during training to prevent benchmark regression.

The operating model is closer to continuous improvement than one-time training. By repeatedly applying reinforcement learning pipelines based on internal evaluation and feedback, models are continuously adjusted in response to regulatory changes, system updates, and new incoming data.

Use cases are clear as well. Government agencies can build models tailored to policy, regulation, and administrative procedures; financial institutions to compliance and risk; software teams to proprietary codebases and development standards; and manufacturers to equipment specifications and maintenance records. The core goal is to operate more accurate and reliable models and agents within an organization's knowledge and constraints.