SAN JOSE, Calif. – The graveyard of enterprise AI projects is littered with failures, rarely due to a lack of processing power, but rather a fundamental understanding gap.
Most large language models (LLMs) are trained on the vast, chaotic expanse of the public internet, leaving them oblivious to the nuanced internal workflows and decades of institutional knowledge that define a specific business.
Mistral AI, the French champion of the European tech scene, believes it has found the solution. On Tuesday at NVIDIA Corp.’s GTC conference here, the company unveiled Mistral Forge, a specialized platform designed to let corporations and governments build custom models trained from scratch on their own proprietary data.
(In December, AWS debuted Nova Forge, which lets enterprises build customized models.)
While competitors like OpenAI and Anthropic have dominated the consumer spotlight, Mistral has quietly fortified its position in the B2B sector. And the strategy appears to be paying off: CEO Arthur Mensch noted the company is on track to surpass $1 billion in annual recurring revenue this year.
“What Forge does is it lets enterprises and governments customize AI models for their specific needs,” said Elisa Salamanca, Mistral’s head of product.
Unlike standard industry practices such as Retrieval-Augmented Generation (RAG) or simple fine-tuning, which essentially layer data on top of an existing brain, Mistral’s Forge allows for more fundamental retraining. This approach promises better performance for non-English languages, highly technical domains, and agentic systems that require reinforcement learning.
The platform leverages Mistral’s library of open-weight models, including the compact Mistral Small 4. Timothée Lacroix, Mistral’s co-founder and chief technology officer, explained that Forge allows companies to optimize smaller, more efficient models for specific tasks.
“The trade-offs we make when we build smaller models is that they cannot be as good on every topic as their larger counterparts,” Lacroix said. “The ability to customize them lets us pick what we emphasize and what we drop.”
In a move reminiscent of Palantir or IBM Corp., Mistral is pairing its software with human expertise. Forge includes access to forward-deployed engineers (FDEs) who embed with clients to clean data and build evaluation frameworks, skills many enterprises currently lack.
The platform has already gained traction with high-profile partners, including Ericsson and the European Space Agency; ASML, the Dutch semiconductor giant that led Mistral’s Series C funding round; and government agencies in Singapore (DSO and HTX).
According to Mistral Chief Revenue Officer Marjorie Janiewicz, the primary targets for Forge remain high-stakes sectors: governments seeking cultural alignment, financial institutions facing strict compliance, and manufacturers requiring deep technical customization.
