Today, we’re introducing Forge, a system for enterprises to build frontier-grade AI models grounded in their proprietary knowledge. 🌎 Forge bridges the gap between generic AI and enterprise-specific needs. Instead of relying on broad, public data, organizations can train models that understand their internal context embedded within systems, workflows,... and policies, aligning AI with their unique operations. We have already partnered with world-leading organizations, like ASML, DSO National Laboratories Singapore, Ericsson, European Space Agency, Home Team Science and Technology Agency (HTX) Singapore and Reply to train models on the proprietary data that powers their most complex systems and future-defining technologies.show more

📖 Learn more here:

Just talked about exactly that at work today. The biggest treasure and advantage Europe has, is its knowledge accumulated over centuries in specialized fields, and the industrial foundation for many things, incl. the skills and knowledge to use it professionally. Noone has that.

Cool, but you need to compete aggressively in the public LLM space and feel the heat, not find yourself a nice cosy niche. That's what Europe needs.

👩‍💻🔥

From public data to proprietary workflows: AI is getting its security clearance. Now imagine these cleared agents handling sensitive cross-border transactions with cryptographic trust. The enterprise agent economy needs rails that match their clearance level.

Enterprise fine-tuning has been possible for a while, but the tooling gap kept it as a consulting project. Making it a product changes the adoption curve. The bet is that enterprise data is differentiated enough to justify the cost of a custom model over prompt engineering on a general one. For finance and healthcare, that's probably true.

This is exactly the direction enterprise AI is moving in. As organizations build frontier-grade models grounded in proprietary knowledge, the next layer becomes critical: runtime trajectory security. That is why I am building ASA 5 - an external runtime security layer for long-horizon AI systems. Forge can help enterprises create models aligned with their internal context. ASA 5 is designed to help monitor whether those systems stay stable over time: - trajectory integrity - drift detection - pre-incident signals - human escalation and guidance Enterprise AI will not only need better knowledge. It will need continuous guidance.

the proprietary data angle is what separates the real enterprise plays from the demos

I bet there's a nice Knowledge Graph and some sort of agent Smith melting all the data sources together in this Forge 🌋

The gap between generic AI and actual business needs is massive. Most enterprise AI projects fail because they're trying to force-fit public models onto proprietary workflows. Training on your own data seems like the obvious solution everyone's been waiting for.

I think systems like Prism AI and Cube AI will be highly needed here for Confidential Computing training and Secure Multi-party Computation

How does this deep customization impact total workflow cost for enterprises compared to advanced fine-tuning or RAG?

Super !

enterprise AI gets real when the model learns the org chart, not just the docs

Impressionnant ! Forge représente une vraie avancée pour les entreprises qui veulent tirer parti de l'IA sans sacrifier la confidentialité de leurs données. Hâte de voir les cas d'usage concrets avec des partenaires comme ASML ou l'ESA. 🚀 Bravo @MistralAI !

Forge finally gives enterprises real control to pre-train + align frontier models directly on proprietary data without the usual RAG. Huge for regulated/defense workflows where context fidelity matters most. 👀

Mistral Forge is a smart move for enterprises who actually have proprietary data worth training on. But here's where it gets interesting: when models are grounded in company-specific knowledge, the safety question shifts from 'is the model safe' to 'how do we trust the data feeding it?' That eval and access control layer isn't just infrastructure anymore - it's the new product differentiator. Anyone else seeing this as the real bottleneck for enterprise adoption?

we need to chat! putting last hand to mvp of nimsforest the organization engine, now routing to 'general' models - with this it will become all powerfull and we can train on the go with new daily data

And another shit

Granite style ? Love your work

Generic models hit a ceiling fast. Proprietary knowledge integration is the real enterprise AI differentiator. This addresses the right problem.

Nuts 🌰 Congrats!

À noter : Forge cible la fine-tuning enterprise sur données propriétaires — segment où Anthropic et OpenAI restent vagues. Mistral capitalise sur le terrain souverain où les DSI européennes ne veulent pas exporter leur knowledge graph chez les hyperscalers US.

😎 Two big AI training announcements dropped today 🍾 Free, open-source (beta), no-code local web UI that makes fine-tuning and running models ridiculously easy and cheap on your own hardware 😁

The data moat is the whole story. A model trained on your internal systems will outperform any generic one on your domain by a lot.

this feels like a practical step for proprietary data. how will updates stay current?

Great!

propietary data is the next step

Secure a Top EU Ai Domain Enhance your online presence. Make it yours: ➡️

Forge lands at the right moment. Reuters said OpenAI is refocusing on coding and business users, so enterprise buyers now care less about generic demos and more about models trained on internal policy and workflow data.

based

curious how this fairs up against @OpenAI’s frontier and other enterprise platforms

Institutional AI is a very necessary step we need to embrace now. Forge looks like it could be very beneficial to institutions that may want to train on their enormous data to develop frontier models. This could in return 'forge' how they make operate to yield outcome.

Grounding on internal workflows is the hard part. Eval and access controls decide everything.