What Is Forge

Update November 2, 2026 — Le Chat is now Vibe. As of May 28, 2026, Le Chat has been rebranded as Vibe (still at chat.mistral.ai): a single agent for work and code that carries over your conversations, settings, and subscriptions. The legacy Le Chat experience remains available in Chat mode; in Work mode, agents are now skills (invoked with @), memories form the knowledge base, and deep search is the Deep Research skill. Plans are now called Mistral Free, Pro, Team, and Enterprise. This lesson uses the name Vibe; note that some menus described here may have changed names or locations.

A model built around your enterprise

Every organization carries a history made of years of decisions, lessons learned, and processes refined by experience. Over time, this history becomes institutional knowledge: how problems are framed in your company, how risks are assessed, how systems are built, and how decisions are made. This knowledge lives in internal documents, technical playbooks, regulatory frameworks, codebases, and even the silent judgments made every workday.

General-purpose models inherit none of this: they know the internet, not your company. Mistral Forge, announced at the NVIDIA GTC conference in March 2026, is Mistral AI’s answer to this gap. It customizes Mistral’s frontier models on your organization’s proprietary data to deliver an AI that reasons like your enterprise.

Two levers that don’t compete

Until now, the primary way to adapt a model to business needs was RAG (Retrieval-Augmented Generation)—exactly what Enterprise Search does. RAG leaves the model unchanged and injects document context with every query; it excels when the question is, “What is the termination clause in last year’s contract?” Fine-tuning operates on a different level: it modifies the model’s weights to permanently encode your knowledge. The request then becomes, “Draft this clause the way our legal department would,” and no document excerpt alone can deliver that. These approaches don’t replace each other; they layer on top of one another.

Forge delivers value across four dimensions. First, the model speaks your company’s language: it knows your internal acronyms refer to specific repositories and stops translating them into generalities. Second, it applies your processes, so an approval request follows your actual sign-off chain rather than a generic workflow. Third, it embeds your expertise—technical diagnostics, risk assessments, sector-specific recommendations—instead of reconstructing it by inference. Finally, it adopts your tone, matching your internal communications, customer interactions, and technical documentation, which cuts down on post-generation rewriting.

Technical architecture

Forge never starts from a blank slate. The starting point is a Mistral frontier model—Mistral Large or Mistral Medium—retaining all its general reasoning, multilingual, and coding capabilities, now enriched with your expertise. You aren’t trading versatility for specialization; you’re adding the latter to the former.

Training runs on Mistral’s secure infrastructure, with guarantees that matter as much as performance: your training data is encrypted and isolated, compute runs on dedicated GPUs in Europe, the resulting model is proprietary and yours to own, and the data used to train it is deleted after delivery. The customized model then deploys according to your constraints—on Mistral infrastructure, where it becomes accessible via Mistral Enterprise and the API; in a dedicated VPC for complete network isolation; on-premise on your own servers for total control; or on a partner cloud (Azure, AWS, or GCP) via existing integrations.

Forward-Deployed Engineers

Forge’s most distinctive element isn’t technical. It’s the dedicated Mistral engineers, called Forward-Deployed Engineers (FDEs), who work directly with your teams. They identify use cases where a customized model truly adds value, structure and clean training datasets, tune hyperparameters to maximize performance, measure quality against benchmarks specific to your industry, and guide integration into your existing systems. This engagement model, inspired by practices at leading technology companies, ensures customization meets your actual needs rather than an approximate interpretation of your brief.

What early partnerships show

Forge launched with partners that illustrate its range of use cases: ASML applies it to lithography optimization for semiconductors, the European Space Agency to processing Earth observation data, Ericsson to telecommunications network optimization, and DSO National Laboratories in Singapore to defense and security applications. Four sectors, four vocabularies, four constraint regimes—and a shared need for performance, security, and sovereignty that a generic model alone cannot satisfy. The next lesson details these deployments.

Key takeaways

• Forge customizes Mistral’s frontier models on your proprietary data
• Fine-tuning complements RAG: one changes model behavior, the other injects context
• Training runs on dedicated GPUs in Europe; your data is deleted after delivery
• You own the model, which deploys on Mistral infrastructure, in a VPC, on-premise, or on a partner cloud
• Forward-Deployed Engineers guide the project from scoping to integration