On March 17, 2026, Mistral launched the Forge platform at NVIDIA GTC. The one-line pitch: enterprises and governments train their own custom models from scratch on their own data — not fine-tuning an existing model, not RAG retrieval, but actual from-the-ground-up training. Head of product Elisa Salamanca puts it plainly: “What Forge does is it lets enterprises and governments customize AI models for their specific needs.”
The route diverges openly from the OpenAI/Anthropic mainstream, whose staples are fine-tuning and RAG — bolting knowledge or behavior onto an untouched model. Mistral is betting the other way.
The Case for Training From Scratch
Per TechCrunch’s analysis, from-scratch training theoretically handles non-English and domain-specific data better, enables reinforcement-learned agentic systems, and reduces dependence on third-party providers — avoiding “risks like model changes or deprecation.” Developers in 2026 feel this one personally: a single provider-side model iteration can shift your application’s behavior wholesale.
Customers build on Mistral’s library of open-weight models, including Mistral Small 4, which shipped alongside Forge. Co-founder Timothée Lacroix’s answer to the small-model ceiling is pragmatic: smaller models “cannot be as good on every topic as their larger counterparts,” and customization “lets us pick what we emphasize and what we drop.”
The FDE Model: Engineers Embedded in the Customer
Forge’s delivery borrows from IBM and Palantir: forward-deployed engineers (FDEs) embed directly with client teams. Salamanca says Forge “already comes with all the tooling and infrastructure so you can generate synthetic data pipelines,” and on evals and data expertise, “that’s what the FDEs bring to the table.” Lacroix adds that Mistral advises on models and infrastructure, but “both decisions stay with the customer.”
The early customer list carries weight: Ericsson, the European Space Agency, Reply, Singapore’s DSO and HTX, and ASML — which led Mistral’s Series C at a roughly $13.8 billion valuation. Target buyers are governments (language and culture tailoring), compliance-heavy finance, manufacturers, and tech companies tuning models to their own codebases.
Who Actually Needs Custom Models?
Futurum Group’s analysis poses the key question: Forge “takes aim at RAG,” but who actually needs custom models? Their answer is restrained — only a small set of organizations with high data maturity, for whom RAG and fine-tuning genuinely fall short, will find from-scratch training worth it. For most enterprises, fine-tuning plus RAG remains the best cost-benefit combination.
That is also what makes the Forge strategy smart: Mistral doesn’t pretend every enterprise should train its own model. It serves, deeply, the narrow slice with the strongest sovereign-AI demand — governments, defense, regulated industries. CEO Arthur Mensch says the enterprise focus is paying off: Mistral is on track to surpass $1 billion in ARR this year, even as OpenAI and Anthropic lead in consumer adoption.
What to Watch
Forge is a snapshot of the AI market’s segmentation: beyond the frontier-model race, “data sovereignty plus customization plus no vendor lock-in” is a real, paying branch. For enterprise technology leaders, the decision order should be: ask whether RAG is enough, then whether fine-tuning is enough, and only then consider training from scratch — and once you reach step three, platforms like Forge start earning their keep.
Sources
- Mistral bets on ‘build-your-own AI’ as it takes on OpenAI, Anthropic — TechCrunch
- Mistral Forge Takes Aim at RAG, But Who Actually Needs Custom Models? — Futurum Group
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.