Mistral bets on ‘build-your-own AI’ in enterprise race
Mistral launches Forge, letting companies train custom AI models on their own data to compete with OpenAI and Anthropic
San Francisco: Most enterprise AI projects fail because companies use models trained on the internet rather than their own business data. Models don’t understand decades of internal documents, workflows, and special knowledge. Mistral, a French AI startup, sees this gap as an opportunity. On Tuesday, the company announced Mistral Forge, a platform that lets businesses build custom AI models using their own data.
Forge lets enterprises and governments customize AI models for their special needs, according to Elisa Salamanca, Mistral’s head of product. Several companies already offer to fine-tune existing models or add company data on top using retrieval augmented generation. These methods don’t retrain models from scratch. Mistral says it allows companies to train models from the beginning.
In theory, this could help with problems that other methods have. For example, better handling of non-English or very specific data, and more control over how models behave. It could also let companies train smarter AI systems and use fewer outside model providers. Forge customers can build models using Mistral’s library of open-weight AI models, including small ones like Mistral Small 4.
Timothée Lacroix, Mistral’s co-founder, explains that smaller models cannot be as good on every topic as bigger ones. The ability to customize them lets Mistral pick what to emphasize and what to drop. Mistral gives advice on which models and setup to use, but customers make the final decisions. Forge also includes Mistral’s team of forward-deployed engineers who work directly with customers.
As a product, Forge comes with tools to create synthetic data pipelines, Salamanca said. But understanding how to build good evaluations and having the right amount of data is hard for most companies. That’s what the forward-deployed engineers provide. Mistral has given Forge to partners like Ericsson, the European Space Agency, Reply, and Singapore’s DSO and HTX. Firms like Ericsson want to adjust models for their needs, financial companies want high compliance, and tech companies need to fit models to their code.