Mistral AI Forge

Mistral AI Forge, introduced by Mistral AI in late 2025, is a comprehensive enterprise model customization platform that enables organizations to build frontier-grade AI systems anchored in their proprietary data, documents, and domain expertise. Unlike lightweight fine-tuning APIs, Forge combines the full lifecycle of model development — continued pretraining, supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and direct preference optimization (DPO) — into a single managed pipeline operated either on Mistral's infrastructure or inside the customer's VPC or on-premises environment.

The platform is built on top of Mistral's open-weight foundation models (Mistral Large 2, Mistral Small 3, Codestral, Pixtral) as well as optional proprietary checkpoints, with typical engagements producing custom models in the 7B to 123B parameter range. Forge handles data curation, tokenization, distributed training across H100/H200 GPU clusters, evaluation harnesses, and deployment — reportedly reducing time-to-production from 9+ months with in-house teams to 6-12 weeks. Mistral reports that Forge customers retain 100% ownership of resulting model weights, a notable contrast with OpenAI's fine-tuning API where derivative weights remain hosted by OpenAI.

Target customers are regulated and data-sensitive industries: banking (BNP Paribas is a named partner), defense, telecoms, healthcare, and legal. Forge is designed for scenarios where sending data to third-party APIs is prohibited, where domain vocabulary is underrepresented in base models, or where sovereignty (EU AI Act, GDPR, sector-specific regulation) requires controlled deployment. It competes with OpenAI's custom model program, Cohere's North, Databricks Mosaic AI, AWS Bedrock Custom Models, and Google Cloud's Vertex AI tuning — differentiating primarily on European data sovereignty, open-weight lineage, and full IP transfer. As of Q2 2026, pricing remains quote-based with engagements starting in the low-to-mid six figures USD.