Mistral AI Unveils Forge for Building Custom Frontier Models on Private Data

Mistral AI has officially unveiled Forge, a comprehensive system designed for enterprises to develop frontier-grade AI models grounded in their own internal knowledge. Announced at the Nvidia GTC conference, this platform allows organizations to move beyond generic chatbots by training models on proprietary codebases, compliance policies, and operational records. Forge provides a path for businesses to conduct full pre-training and reinforcement learning within their own infrastructure.

The system stands apart from traditional AI offerings by supporting both dense and mixture-of-experts architectures. This flexibility ensures that companies can optimize their custom models for specific performance needs or compute costs. By integrating multimodal capabilities, Forge allows these models to process text, images, and structured data simultaneously to capture the full context of an organization.

A core component of the launch is the agent-first design, which integrates deeply with autonomous tools like Mistral Vibe. Autonomous agents can now manage complex machine learning workflows, including hyperparameter optimization and synthetic data generation, using plain English commands. This reduces the need for massive internal research teams while increasing the reliability of enterprise agents in real-world environments.

The initial rollout includes partnerships with world-leading organizations such as ASML, Ericsson, and the European Space Agency. These partners use Forge to internalize decades of institutional decisions and engineering standards into their AI systems. Agencies in Singapore, including HTX and DSO National Laboratories, are also leveraging the platform for highly sensitive governmental applications where data security is paramount.

While services like ChatGPT offer generic reasoning through a rental model, Forge shifts the focus toward strategic autonomy and model ownership. Organizations maintain absolute control over their training data and the final model weights, ensuring that sensitive intellectual property never leaves their environment. This approach transforms AI from an external tool into a core, evolving piece of corporate infrastructure.