Mistral Forge, a new system from Mistral AI, enables enterprises to build artificial intelligence models trained on their own proprietary data rather than generic public datasets.
The company said Forge addresses a gap between general-purpose AI and the specific operational needs of large organizations. Enterprises rely on internal knowledge — engineering standards, compliance policies, codebases, and institutional decisions — that public AI models do not capture.
How Mistral Forge Works
Mistral Forge supports multiple stages of the model lifecycle. Organizations can use pre-training to build domain-aware models from large internal datasets. Post-training methods allow teams to refine model behavior for specific tasks. Reinforcement learning helps align models with internal policies and operational objectives.
The system also supports both dense and mixture-of-experts (MoE) model architectures. Dense models offer broad capability across enterprise tasks, while MoE enables large models to run with lower latency and reduced compute cost. Forge additionally supports multimodal inputs, allowing models to learn from text, images, and other data formats.
Enterprise Control and Data Governance
Mistral AI said Forge gives organizations direct control over how their knowledge is encoded into AI systems. Models can be trained using proprietary datasets and governed by internal policies, evaluation standards, and compliance requirements. This is particularly relevant for regulated industries where governance frameworks are mandatory.
By operating within their own infrastructure environments, enterprises retain ownership of the resulting models. The company positioned this as a form of strategic autonomy as AI becomes embedded in core business operations.
“Forge enables enterprises to build and continuously improve models trained on their own data and aligned with their operational context.”
Mistral AI
Agent-First Design and Continuous Improvement
Forge is designed with autonomous agents as primary users. The company said an agent such as Mistral Vibe can use Forge to fine-tune models, identify optimal hyperparameters, schedule training jobs, and generate synthetic data. Forge monitors metrics throughout the process to prevent model regression on key benchmarks.
The system also supports continuous improvement through reinforcement learning pipelines. As enterprise environments evolve — with regulatory changes, system updates, and new data — organizations can refine model behavior using feedback from internal evaluations and operational workflows.
Early Partners and Enterprise Applications
Mistral AI said it has already worked with several organizations to train models on proprietary data. Partners include ASML, DSO National Laboratories Singapore, Ericsson, the European Space Agency, Home Team Science and Technology Agency (HTX) Singapore, and Reply.
The company outlined several application areas for Forge. Government agencies can train models on policy frameworks, regulatory texts, and administrative procedures. Financial institutions can apply Forge to compliance frameworks and risk documentation. Software teams can train models on proprietary codebases and development standards to improve output quality across the software development lifecycle. Manufacturers can use Forge with engineering specifications and maintenance records to support diagnostics and operational decision-making.
Outlook
Mistral AI said AI models are becoming a foundational layer of enterprise infrastructure. As organizations integrate AI agents into core operations, the ability to encode institutional knowledge into model behavior will grow in importance. The company said Forge is designed to allow organizations to treat AI models as strategic assets that evolve alongside their knowledge, processes, and expertise. Enterprises interested in the platform can sign up through the Mistral AI website to learn more.