French artificial intelligence startup Mistral AI has launched a new platform called Forge, designed to enable enterprises to build custom AI models trained exclusively on their proprietary data. The announcement was made at Nvidia's GTC conference in San Francisco on March 18, 2025, and represents a direct challenge to established enterprise AI players like OpenAI and Anthropic.
The platform addresses a critical gap in enterprise AI adoption, where generic models trained on public internet data often fail to understand specific business contexts, internal documents, workflows, and institutional knowledge. "What Forge does is it lets enterprises and governments customize AI models for their specific needs," explained Elisa Salamanca, Mistral's head of product.
Unlike competitors that offer fine-tuning or retrieval augmented generation (RAG) techniques, Forge enables true from-scratch model training. This approach offers advantages in domain-specific optimization (including non-English languages), greater behavioral control over model outputs, reduced third-party dependency, and the ability to develop agentic systems using reinforcement learning.
Mistral's enterprise focus is already yielding significant financial results. CEO Arthur Mensch revealed the company projects surpassing $1 billion in annual recurring revenue this year. The company's valuation reached €11.7 billion during its Series C funding round in September 2024, led by semiconductor equipment manufacturer ASML, which is both an investor and a customer.
Early adopters of Forge include major corporations and government agencies such as telecommunications giant Ericsson, the European Space Agency, Italian consulting firm Reply, Singapore's DSO and HTX agencies, and ASML. According to Marjorie Janiewicz, Mistral's chief revenue officer, primary use cases span government agencies needing cultural/linguistic tailoring, financial institutions with high-compliance requirements, manufacturing companies seeking production optimization, and technology firms requiring codebase-specific tuning.
Mistral employs a forward-deployed engineering model, where engineers embed directly with customer teams to identify relevant data sources and adapt systems to specific operational needs. "Understanding how to build proper evaluations and ensuring sufficient data quality requires specialized knowledge," Salamanca emphasized.
The launch occurs during intense competition in enterprise AI, with OpenAI's GPT-4 Enterprise and Anthropic's Claude for Business dominating recent conversations. Mistral differentiates itself through its open-weight model philosophy and emphasis on customization and data control, appealing to organizations with stringent data protection requirements, particularly in European markets.
