Mistral AI Launches Forge to Let Companies Train Custom AI on Their Own Data

Many companies desire their own AI model on in-house servers, but putting that into practice is no easy feat. This is where Mistral AI, currently Europe’s most valuable AI company, steps in. The French unicorn has unveiled a platform called Forge that enables companies to train their own large language models based on Mistral’s open-weight models. In doing so, the startup is addressing a central criticism of common enterprise AI solutions: most available models were trained on publicly accessible data and understand neither the internal terminology nor the specific processes of a given company.

Open Weight vs. Open Source: An Important Distinction

Before understanding what Forge does, it is worth examining a term that is frequently misunderstood in the AI industry: Open Weight. Mistral releases its models as open-weight models, meaning that the trained model weights — the numerical parameters that determine the model’s behavior — are publicly accessible and can be downloaded.

This sounds like open source, but it is not necessarily so. The difference lies in the details:

With open-weight models, developers and companies gain access to the fully trained model. They can run it locally, adapt it, and develop it further, without knowing on exactly which data it was originally trained or what the complete training process looked like. Open source, by contrast, means in the classical sense that the entire source code, training data, and methodology are disclosed, so that the model can be fully reproduced.

What Forge Specifically Enables

Forge builds on these open-weight models and gives companies the tools to fundamentally continue training them with their own data. This distinguishes the approach from common alternatives such as Retrieval Augmented Generation (RAG) or simple fine-tuning, in which the base model itself remains unchanged and company data is merely consulted at runtime.

Mistral describes three training stages that Forge supports:

• Pre-Training: Companies can train models from scratch with large internal datasets, allowing the model to deeply internalize domain-specific knowledge.
• Post-Training: Existing models are refined for specific tasks and environments, such as particular workflows or technical terminology.
• Reinforcement Learning: Models are aligned according to internal guidelines, evaluation criteria, and operational objectives, which is particularly relevant for use in autonomous agent systems.

“What Forge does is enable companies and governments to adapt AI models to their specific needs.” (Elisa Salamanca, Head of Product at Mistral)

Control Over Data and Models Remains with the Company

A central promise of Forge is strategic autonomy. Companies train their models on their own data, operate them within their own infrastructure, and retain full control over how their institutional knowledge flows into the model’s behavior. This is particularly relevant for regulated industries in which compliance requirements and internal governance frameworks must be adhered to.

Mistral co-founder Timothée Lacroix explains the practical advantage of smaller, customized models over large general-purpose models as follows:

“The trade-offs we make when building smaller models mean they cannot be as good as their larger counterparts on every topic. The ability to customize them allows us to choose what we emphasize and what we leave out.”

Technical Flexibility: Dense and MoE Architectures

Forge supports two fundamental model architectures between which companies can choose depending on their requirements:

• Dense models offer strong general performance for a broad range of enterprise tasks.
• Mixture-of-Experts (MoE) models enable very large models to be operated more efficiently, with lower latency and reduced computational costs at comparable performance.

In addition, Forge supports multimodal inputs, allowing models to learn not only from text but also from images and other data formats.

Use Cases in Practice

Mistral has already tested Forge with a number of partner companies and government agencies, including ASML, Ericsson, the European Space Agency (ESA), the Italian consulting firm Reply, and the Singaporean agencies DSO National Laboratories and HTX. The platform was unveiled at the Nvidia GTC conference.

Typical areas of application cited by Mistral include:

• Government agencies and administrations: Models tailored to specific languages, dialects, legal frameworks, and administrative procedures.
• Financial institutions: Models that have internalized compliance requirements, risk procedures, and regulatory documentation.
• Software development teams: Models trained on internal codebases and development standards to provide context-sensitive support throughout the entire development cycle.
• Manufacturing companies: Models for diagnostics, design analysis, and operational decision-making based on specifications and maintenance logs.

Accompanying Expertise Through Embedded Engineers

In addition to the technical platform, Mistral offers so-called Forward-Deployed Engineers (FDEs) who are deployed directly at client sites. These specialists help identify the right data, build evaluation frameworks, and optimize training pipelines. Salamanca emphasizes that the knowledge of how to define the right evaluation metrics and compile sufficiently high-quality data is still lacking in most companies.

Mistral, which most recently raised capital at a valuation of around €11.7 billion and claims to be on track for more than €1 billion in annual recurring revenue, is positioning itself with Forge clearly as a provider for enterprise customers seeking greater control over their AI systems than pure cloud services have so far allowed.