Mistral AI launched Forge at NVIDIA GTC in San Jose on March 17, a platform that lets enterprises train frontier-grade AI models from scratch using their own proprietary data, with NVIDIA GPU optimization and forward-deployed engineering support.

What Happened

Mistral announced Forge at NVIDIA GTC 2026, positioning it as a full-stack platform for enterprises that want to build custom AI models rather than fine-tune existing ones. Unlike most enterprise AI offerings that start from a pretrained base model, Forge enables training from the ground up on a company's own data.

The platform includes an open-weight model library (featuring Mistral Small 4, released the same week), data pipeline tools covering data acquisition, curation, and synthetic data generation, and forward-deployed engineers who embed directly with customer teams to guide model development.

Forge supports multi-environment deployment across cloud and on-premises infrastructure with no vendor lock-in, and is optimized for NVIDIA's latest GPU architectures. Early partners include ASML, DSO National Laboratories Singapore, the European Space Agency, and Reply.

Why It Matters

Most enterprise AI platforms offer fine-tuning or retrieval-augmented generation on top of existing foundation models. Forge takes a different approach by letting companies train models entirely from their own data, which matters for industries with highly specialized knowledge, sensitive data, or regulatory requirements that make shared foundation models impractical.

The forward-deployed engineering model is notable. Rather than selling software and leaving customers to figure out implementation, Mistral embeds its own engineers with partner teams. This approach mirrors what defense contractors and enterprise database companies have done for decades, applied here to AI model development.

CEO Arthur Mensch stated that Mistral is on track to surpass $1 billion in annual recurring revenue this year. That figure, combined with Forge's enterprise positioning, signals Mistral is competing directly with OpenAI and Anthropic for large enterprise contracts, differentiating on openness and customization rather than model scale alone.

Key Details

  • Training: Full from-scratch model training on proprietary enterprise data, not just fine-tuning.
  • Model library: Open-weight models including Mistral Small 4 as starting points or reference architectures.
  • Data tools: Built-in data acquisition, curation, and synthetic data generation pipelines.
  • Deployment: Multi-environment support across cloud and on-premises, no vendor lock-in.
  • Hardware: Optimized for NVIDIA's latest GPU architectures.
  • Support: Forward-deployed engineers embedded with customer teams.
  • Early partners: ASML, DSO National Laboratories Singapore, European Space Agency, Reply.
  • Related releases: Mistral Small 4, Leanstral (open-source code verification agent), and NVIDIA Nemotron Coalition membership announced the same week.

What to Do Next

Enterprises evaluating custom AI model development should review Mistral's Forge documentation to understand whether training from scratch fits their use case better than fine-tuning existing models. For teams already using Mistral's API or open-weight models, Forge represents a natural next step when off-the-shelf models cannot capture domain-specific knowledge. Compare Forge's from-scratch training approach against fine-tuning offerings from other providers to determine which delivers better results for your specific data and requirements.