Mistral AI, the Paris-based artificial intelligence company, has unveiled Forge, a platform that allows organisations to train large AI models from scratch on their own documents, code and operational records, keeping the resulting models entirely within their own infrastructure.
The product addresses a problem that has become increasingly visible as enterprises try to move beyond general-purpose AI tools: publicly trained models know a great deal about the world but relatively little about the internal standards, compliance requirements and operational procedures that govern how any specific organisation actually functions.
Forge is designed to close that gap by embedding institutional knowledge directly into model behaviour rather than relying on prompts or retrieval systems to supply context at runtime.
The platform supports the full training lifecycle, covering pre-training on raw data, post-training refinement to adjust model behaviour, and reinforcement learning to align outputs with internal evaluation criteria and policy constraints.
It is built to handle both dense transformer architectures and mixture-of-experts (MoE) models, a design approach that activates only a subset of a model's parameters for any given task, making large models more computationally efficient to run.
Multimodal inputs are supported, meaning organisations can train on a combination of text, code, images and structured data rather than text alone.
The platform is also designed to run training and evaluation pipelines on a continuous basis rather than as periodic projects, which Mistral positions as essential for organisations where internal standards and procedures evolve regularly.
Mistral has disclosed early partnerships with ASML, the Dutch semiconductor equipment maker, Ericsson, the European Space Agency, and defence and security agencies in Singapore, suggesting the initial commercial focus is on technically complex, regulated environments where data sovereignty is a hard requirement rather than a preference.
The agent-first framing is also significant
Related reading
- AuraLift AI launches round-the-clock mental wellness coaching platform aimed at adults who avoid traditional…
- Qurrent raises $15m to replace back-office workers with AI agents
- Nordic Semiconductor expands 'ultra low power' AI chip lineup
Mistral's own Vibe agent can manage hyperparameter tuning, job scheduling and synthetic data generation within Forge, reducing the engineering burden on in-house teams and allowing smaller technical organisations to run sophisticated training workloads without large machine learning infrastructure teams.
The competitive context is notable
Forge positions Mistral directly against the cloud hyperscalers and larger AI labs that offer fine-tuning services on their own infrastructure, with the differentiating argument that organisations using Forge retain complete control over both the training data and the resulting model weights, a distinction that is increasingly commercially meaningful in financial services, defence and healthcare.
The recap
- Mistral AI launches Forge to train models on proprietary data.
- Supports dense and mixture-of-experts architectures and multimodal inputs.
- Organizations can train, align and evaluate models using Forge.
