French AI startup Mistral is making an aggressive play for the heart of the enterprise market with a new platform that lets companies build their own AI models from the ground up, directly challenging US heavyweights OpenAI and Anthropic on their most lucrative turf: corporate AI deployments.
Unveiled on Tuesday at Nvidia’s GTC 2026 conference in San Jose, the platform, called Mistral Forge, promises to let enterprises and governments train custom models from scratch on their own proprietary data, rather than merely fine‑tuning general-purpose systems trained on the public internet.
A direct shot at the enterprise AI status quo
For years, most enterprise AI offerings have revolved around adapting existing foundation models using techniques like fine-tuning or retrieval augmented generation (RAG), where company data is layered on top and queried at runtime instead of being baked into the model itself.
Mistral is openly positioning Forge as a break from that playbook. “What Forge does is it lets enterprises and governments customize AI models for their specific needs,” Elisa Salamanca, Mistral’s head of product, told TechCrunch in an interview. “The models that are trained on the open internet don’t always understand your organization’s language, edge cases or workflows, and that’s where we see a lot of AI projects fail,” she added.
The company argues that by allowing organizations to train models from scratch on domain-specific and often non‑public data, Forge can deliver better performance on highly specialized tasks, especially in non‑English markets, regulated industries and complex industrial environments.
‘Build-your-own AI’ as a business thesis
Mistral’s bet is that a growing share of large enterprises will want not just access to powerful models, but genuine ownership and deep control over them from the training corpus to the evaluation metrics and deployment footprint.
“Many AI initiatives inside enterprises fail not because the technology is missing, but because the models don’t really understand the business,” CEO Arthur Mensch has argued, pointing out that most frontier models are trained on generic internet data rather than decades of internal documentation and institutional knowledge. Mensch says Mistral’s strategy of focusing squarely on corporate clients is paying off, with the company on track to surpass 1 billion dollars in annual recurring revenue this year.
In official materials announcing the launch, Mistral described Forge as “a system that allows enterprises to build frontier-grade AI models grounded in their proprietary knowledge,” framing it as a way to convert internal data assets into durable AI advantages rather than simply plugging into third‑party APIs.
How Forge works under the hood
Forge is built around Mistral’s library of open‑weight models, including smaller systems such as the recently introduced Mistral Small 4, which are designed to be efficient enough for targeted enterprise deployments but flexible enough to be re‑trained for narrow domains.
“Forge already comes with all the tooling and infrastructure so you can generate synthetic data pipelines,” Salamanca said, noting that the service is designed to support everything from data preparation and training runs to evaluation and monitoring. “But understanding how to build the right evals and making sure that you have the right amount of data is something that enterprises usually don’t have the right expertise for, and that’s what the FDEs bring to the table.”
Those “FDEs” Mistral’s field deployment engineers will work alongside customer teams to select training data, design evaluation suites and tailor models’ behavior, a strategy reminiscent of the high‑touch consulting models used by established enterprise players like IBM and Palantir.
Timothée Lacroix, Mistral’s co‑founder and chief technologist, said the platform is also a way to unlock more value from the company’s existing model family. “The compromises we make when creating smaller models mean they cannot excel in every area as their larger counterparts do, so the ability to customize them allows us to prioritize what we focus on and what we set aside,” he explained.
While Mistral provides guidance on architecture, infrastructure and training strategy, Lacroix stressed that final design and deployment choices rest with customers, including whether to run models on‑premises, in private clouds or in hybrid setups.
Nvidia alliance and the GTC backdrop
The launch of Forge was tightly choreographed with Nvidia’s GTC, which this year has leaned heavily into AI agents, enterprise AI stacks and custom model-building as key themes.
Mistral and Nvidia recently announced a broader strategic partnership to co‑develop open frontier models and to optimize Mistral’s systems for Nvidia’s latest hardware platforms. As part of that collaboration, Forge is designed to run efficiently on Nvidia GPUs, and the companies have been jointly promoting reference architectures for enterprises that want to stand up their own AI training and inference clusters.
At GTC, Mistral executives joined Nvidia sessions focused on “how to build your custom AI advantage,” underscoring the shared message that organizations should be thinking not just about consuming AI, but about cultivating AI capabilities as core infrastructure.
Early adopters and target sectors
Mistral has already rolled out Forge to an initial group of partners and customers, spanning telecoms, government, aerospace, manufacturing and high‑tech.
Among the named early adopters are Ericsson, the European Space Agency, Italian consulting firm Reply, and Singaporean government agencies DSO and HTX. Dutch chipmaking giant ASML which led Mistral’s Series C round last year at a valuation of 11.7 billion euros is also an early Forge customer, using the platform to explore domain‑specific models for complex semiconductor manufacturing workflows.
According to Mistral’s chief revenue officer Marjorie Janiewicz, these pilots reflect the kinds of use cases the company expects to dominate Forge adoption: governments tailoring models to local languages and cultures; financial institutions facing stringent compliance rules; manufacturers with highly customized processes; and technology firms tuning models to their proprietary codebases.
“These are organizations that can’t afford generic behavior from their AI systems,” Janiewicz has said in briefings. “They need systems that understand their regulatory environment, their safety constraints, their domain jargon and they want to know that if a model changes, it’s because they changed it, not because a vendor updated an API.”
Going beyond OpenAI and Anthropic’s fine-tuning model
Although Mistral does not name specific rivals in its launch materials, Forge is clearly positioned against the API‑centric strategies of OpenAI and Anthropic, which primarily offer access to pre‑trained models like GPT‑4‑class systems and Claude through hosted endpoints with fine‑tuning and RAG layered on top.
Industry analysts note that those platforms have been highly successful in getting enterprises started with generative AI, but they stop short of giving customers full control over model training or weights, especially for frontier‑grade systems.
Mistral’s counter‑proposal is that enterprises should be able to train their own proprietary models, with full visibility into the training data, architecture and evaluation criteria even if they leverage Mistral’s open‑weight families as starting points or templates.
In theory, that approach could help organizations reduce dependence on third‑party providers, mitigate the risk of sudden model changes or deprecations, and better align AI behavior with internal policies and risk frameworks. It could also be attractive to non‑US governments and corporations that are wary of relying exclusively on American AI vendors for critical infrastructure.
Technical and organizational hurdles remain
Building custom models from scratch, however, is far from trivial. It demands substantial compute resources, high‑quality labeled data and specialized machine learning expertise, all of which can quickly become bottlenecks for even large enterprises.
Mistral is attempting to lower these barriers in several ways: by providing pre‑built tooling for data pipelines, evaluation and monitoring; by making its open‑weight models available as scaffolds; and by embedding its own experts with customer teams through the field deployment engineer program.
Still, the company acknowledges that not every organization will want or need to go all the way to fully bespoke models. Forge, according to Salamanca, is meant to support a spectrum of options, from conventional fine‑tuning on top of Mistral’s models to complete retraining on organization‑specific corpora. “It’s not about saying everyone must train from scratch,” she said. “It’s about saying that if you need that level of control, it should be possible.”
A high-stakes bet in a crowded field
The launch of Forge comes as the enterprise AI market intensifies, with cloud hyperscalers, model labs and consulting giants all racing to lock in long‑term corporate relationships. OpenAI continues to expand its enterprise and ChatGPT for Business offerings, while Anthropic has been pushing deeper into regulated sectors with its Claude models and safety‑focused tooling.
By contrast, Mistral has from the outset framed itself as an “enterprise‑first” and “open‑weight” alternative, betting that transparency and controllability will matter more over time than pure model size or benchmark scores.
“Mistral is making a decisive pitch to corporate buyers with Forge, a new platform built to let enterprises train and operate their own AI models on proprietary data,” one industry report noted, adding that the initiative “puts the French startup head‑to‑head with OpenAI and Anthropic by betting that ‘build‑your‑own AI’ will beat out generic models for mission‑critical work.”
Whether that thesis plays out will depend on how many organizations are willing to invest in the data, infrastructure and talent required to truly own their AI stack and whether regulators and boards come to see bespoke models as a safer, more controllable alternative to black‑box APIs.
For now, with Forge, Mistral has made its move: offering enterprises not just another powerful model to call, but a toolkit to build models of their own, on their own terms.
Comments