Mistral AI launches Forge to help companies build proprietary AI models, challenging cloud giants

Mistral AI on Monday launched Forge, an enterprise model training platform that allows organizations to build, customize, and continuously improve AI models using their own proprietary data — a move that positions the French AI lab squarely against the hyperscale cloud providers in one of the most consequential and least understood markets in enterprise technology.

The announcement caps a remarkably aggressive week for Mistral, which also released its Mistral Small 4 model, unveiled Leanstral — an open-source code agent for formal verification — and joined the newly formed Nvidia Nemotron Coalition as a co-developer of the coalition's first open frontier base model. Together, these moves paint the picture of a company that is no longer content to compete on model benchmarks alone and is instead racing to become the infrastructure backbone for organizations that want to own their AI rather than rent it.

Forge goes significantly beyond the fine-tuning APIs that Mistral and its competitors have offered for the past year. The platform supports the full model training lifecycle: pre-training on large internal datasets, post-training through supervised fine-tuning, DPO, and ODPO, and — critically — reinforcement learning pipelines designed to align models with internal policies, evaluation criteria, and operational objectives over time.

"Forge is Mistral's model training platform," said Maliena Guy, head of product at Mistral AI, in an exclusive interview with VentureBeat ahead of the launch. "We've been building this out behind the scenes with our AI scientists. What Forge actually brings to the table is that it lets enterprises and governments customize AI models for their specific needs."

Why Mistral says fine-tuning APIs are no longer enough for serious enterprise AI

The distinction Mistral is drawing — between lightweight fine-tuning and full-cycle model training — is central to understanding why Forge exists and whom it serves.

For the past two years, most enterprise AI adoption has followed a familiar pattern: companies select a general-purpose model from OpenAI, Anthropic, Google, or an open-source provider, then apply fine-tuning through a cloud API to adjust the model's behavior for a narrow set of tasks. This approach works well for proof-of-concept deployments and many production use cases. But Guy argues that it fundamentally plateaus when organizations try to solve their hardest problems.

"We had a fine-tuning API relying on supervised fine-tuning. I think it was kind of what was the standard a couple of months ago," Guy told VentureBeat. "It gets you to a proof-of-concept state. Whenever you actually want to have the performance that you're targeting, you need to go beyond. AI scientists today are not using fine-tuning APIs. They're using much more advanced tools, and that's what Forge is bringing to the table."

What Forge packages, in Guy's telling, is the training methodology that Mistral's own AI scientists use internally to build the company's flagship models — including data mixing strategies, data generation pipelines, distributed computing optimizations, and battle-tested training recipes. She drew a sharp line between Forge and the open-source tools and community tutorials that are freely available today.

"There's no platform out there that provides you real-world training recipes that work," Guy said. "Other open-source repositories or other tools can give you generic configurations or community tutorials, but they don't give you the recipe that's been validated — that we've been doing for all of our flagship models today."

From ancient manuscripts to hedge fund quant languages, early customers reveal what off-the-shelf AI can't do

The obvious question facing any product like Forge is demand. In a market where GPT-5, Claude, Gemini, and a growing fleet of open-source models can handle an enormous range of tasks, why would an enterprise invest the time, compute, and expertise required to train its own model from scratch?

Guy acknowledged the question head-on but argued that the need emerges quickly once companies move beyond generic use cases. "A lot of the existing models can get you very far," she said. "But when you're looking at what's going to make you competitive compared to your competition — everyone can adopt and use the models that are out there. When you want to go a step beyond that, you actually need to create your own models. You need to leverage your proprietary information."

The real-world examples she cited illustrate the edges of the current model ecosystem. In one case, Mistral worked with a public institution that had ancient manuscripts with missing text from damaged sections. "The models that were available were not able to do this because they've never seen the data," Guy explained. "Digitization was not very good. There were some unique patterns and characters, and so we actually created a model for them to fill in the spans. This is now used by their researchers, and it's accelerating their publication and understanding of these documents."

In another engagement, Mistral partnered with Ericsson to customize its Codestral model for legacy-to-modern code translation. Ericsson, Guy said, has built up half a decade of proprietary knowledge around an internal calling language — a codebase so specialized that no off-the-shelf model has ever encountered it. "The concrete impact is like turning a year-long manual migration process, where each engineer needs six months of onboarding, to something that's really more scalable and faster," she said.

Perhaps the most telling example involves hedge funds. Guy described working with financial firms to customize models for proprietary quantitative languages — the kind of deeply guarded intellectual property that these firms keep on-premises and never expose to cloud-hosted AI services. Using Forge's reinforcement learning capabilities, Mistral helped one hedge fund develop custom benchmarks and then trained the model to outperform on them, producing what Guy called "a unique model that was able to give them the competitive edge that was needed."

How Forge makes money: license fees, data pipelines, and embedded AI scientists

Forge's business model reflects the complexity of enterprise model training. According to Guy, it operates across several revenue streams. For customers who run training jobs on their own GPU clusters — a common requirement in highly regulated or IP-sensitive industries — Mistral does not charge for compute. Instead, the company charges a license fee for the Forge platform itself, along with optional fees for data pipeline services and what Mistral calls "forward-deployed scientists" — embedded AI researchers who work alongside the customer's team.

"No competitor out there today is kind of selling this embedded scientist as part of their training platform offering," Guy said.

This model has clear echoes of Palantir's early playbook, where forward-deployed engineers served as the critical bridge between powerful software and the messy reality of enterprise data. It also suggests that Mistral recognizes a fundamental truth about the current state of enterprise AI: the technology alone is not enough. Most organizations lack the internal expertise to design effective training recipes, curate data at scale, or navigate the treacherous optimization landscape of distributed GPU training.

The infrastructure itself is flexible. Training can happen on Mistral's own clusters, on Mistral Compute (the company's dedicated infrastructure offering), or entirely on-premises within the customer's own data centers. "We have all these different cases, and we support everything," Guy said.

Keeping proprietary data off the cloud is Forge's sharpest selling point

One of the sharpest points of differentiation Mistral is pressing with Forge is data privacy. When customers train on their own infrastructure, Guy emphasized that Mistral never sees or accesses the customer's data.