Mistral AI Launches Forge at GTC 2026 — Full-Cycle Enterprise Model Training Platform Goes Beyond Fine-Tuning with Pre-Training, RL, and Agent-First Design

Mistral AI announced Forge at NVIDIA GTC 2026 on March 17, positioning it as a comprehensive enterprise model training platform that goes far beyond the fine-tuning APIs offered by OpenAI, Anthropic, and Google.

WHAT FORGE DOES:

Forge supports the complete model training lifecycle:

• Pre-training on large internal datasets to build domain-aware foundation models
• Post-training through supervised fine-tuning, DPO, and ODPO methods
• Reinforcement learning pipelines to align models with internal policies, evaluation criteria, and operational objectives

This is a fundamental shift from the industry-standard approach of fine-tuning general-purpose models via API. Elisa Salamanca, Mistral head of product, told VentureBeat: "AI scientists today are not using fine-tuning APIs. They are using much more advanced tools, and that is what Forge is bringing to the table."

ENTERPRISE PARTNERS:

Mistral has already deployed Forge with major organizations:

• ASML (semiconductor lithography)
• European Space Agency (ESA)
• DSO National Laboratories Singapore
• Home Team Science and Technology Agency (HTX) Singapore
• Ericsson
• Reply

These organizations are training models on proprietary data including engineering standards, compliance policies, codebases, and operational processes.

AGENT-FIRST DESIGN:

Critically, Forge is designed agent-first. Code agents like Mistral Vibe can autonomously:

• Fine-tune models
• Find optimal hyperparameters
• Schedule training jobs
• Generate synthetic data
• Hill-climb evaluations

This positions Forge as infrastructure where agents train other agents, a recursive improvement loop.

ARCHITECTURE FLEXIBILITY:

Forge supports both dense and Mixture-of-Experts (MoE) architectures. MoE enables large models to run efficiently by activating only a subset of parameters per query (e.g., Mistral Small 4 has 119B parameters but only 6B active per query). Multimodal inputs (text, images) are also supported.

STRATEGIC SIGNIFICANCE:

Forge challenges hyperscale providers by offering enterprises ownership and control over their AI models rather than API-only access. In regulated industries, this control is essential for compliance. The platform represents Mistral evolving from a model company to an enterprise AI infrastructure company, competing directly with OpenAI and Google on enterprise training capabilities.

Salamanca argued: "Everyone can adopt and use the models that are out there. When you want to go a step beyond that, you need to create your own models. You need to leverage your proprietary information."

COMPETITIVE CONTEXT:

This caps a major GTC week for Mistral: Small 4 model release, Leanstral open-source code agent, Nemotron Coalition membership with NVIDIA, and now Forge. Together, these moves position Mistral as a full-stack enterprise AI infrastructure provider, not just a model lab.