Most corporate artificial intelligence projects do not fail because the technology does not work. They fail because the model simply does not understand what the company does. Trained on internet data, it does not know your internal processes, corporate terminology, or decades of accumulated decisions. And it is precisely this gap that French startup Mistral AI has decided to fill.
Mistral Forge
At the Nvidia GTC conference, Mistral unveiled a platform called Mistral Forge. It is a system that allows companies to train their own AI models directly on their internal data. No general-purpose models trained on public sources. Instead, a model that knows your regulatory documents, your code, your operational records, and your business processes.
It sounds simple, but in practice, it is quite a bold move. Most competitors offer only fine-tuning of existing models or a technique called RAG (retrieval-augmented generation), in which the model accesses a corporate database when processing a query. Mistral goes further: it enables models to be trained from scratch on proprietary data. This delivers better results, particularly in specialized industries, less common languages, or highly regulated environments.
Mistral has built its reputation primarily among corporate customers, while OpenAI and Anthropic dominate the consumer market. CEO Arthur Mensch acknowledged that the focus on businesses is paying off: the company is aiming for more than one billion dollars in annual revenue. Forge is the logical next step. Companies want control over their data and over how their AI behaves. And Mistral gives it to them.
Elisa Salamanca, Mistral's head of product, summed it up succinctly: "Forge enables companies and governments to adapt AI models to their specific needs." No compromises, no sharing data with third parties.
How does Forge work?
The platform supports the model's entire lifecycle. Companies can begin with pre-training on large internal datasets, continue by fine-tuning behavior for specific tasks, and finally use reinforcement learning to ensure that the model follows internal policies and operational goals. Forge supports both dense models and the mixture-of-experts (MoE) architecture, which can run large models more efficiently with lower latency. It also works with multimodal inputs, including text, images, and other data formats.
An interesting detail: Forge is designed to be used by autonomous agents as well. Mistral's agent Mistral Vibe can use it to fine-tune models, find optimal hyperparameters, or generate synthetic data. The entire process can be launched with a simple English command.
Forge is already being tested by global organizations such as ASML, Ericsson, the European Space Agency, Singapore's DSO National Laboratories, and the Italian consulting firm Reply. ASML is an especially interesting case: the Dutch chip manufacturer led Mistral's funding round last September, when the company reached a valuation of more than 11 billion euros.
These partnerships show where Mistral sees its main customers. They include governments that need models for specific languages and cultural contexts. Financial institutions with strict compliance requirements. Manufacturers that need specialized expertise. And technology companies that want a model perfectly aligned with their code.
Dedicated engineers working directly with customers
Forge does not come as software alone. Mistral also provides a team of so-called forward-deployed engineers, its own engineers who work directly with customers. They help select the right data, establish evaluation criteria, and tailor the model to specific needs. It is an approach Mistral has borrowed from companies such as IBM and Palantir.
Timothée Lacroix, Mistral's co-founder and chief technology officer, explained the reasoning behind it: "The trade-offs we make when building smaller models mean they cannot be equally good at everything. The ability to customize them allows us to choose what to emphasize and what to leave aside."
This is a question the entire industry is now asking. Training models from scratch is expensive and time-consuming. But Mistral argues that the result is worth it: a model that truly understands your company behaves more reliably, selects tools more accurately, and handles complex multi-step processes better. Companies that deploy Forge will gain models as strategic assets, not merely as external tools. Models that evolve alongside the company, adapt to new regulations, and grow with corporate knowledge. That is a different league from a chatbot trained on Wikipedia.
Mistral Forge arrives at a time when enterprise AI is moving from experimentation to real-world deployment. And the French startup is betting that companies will want AI that is truly their own. Time will tell whether it is right. But the first partners suggest that it is not a bad bet.
Additional source: techcrunch.com
