Mistral's Custom AI for Enterprise: Challenging OpenAI and Anthropic

Many enterprise AI initiatives falter, not due to a deficit in technological capability, but because the underlying models frequently lack a deep comprehension of the specific business context. These models are typically trained on vast internet datasets, rather than leveraging an organization's extensive internal documentation, proprietary workflows, and accumulated institutional knowledge.

This critical disparity presents a significant opportunity for Mistral, the French AI startup. The company unveiled Mistral Forge on Tuesday, a new platform empowering enterprises to construct bespoke AI models, meticulously trained on their proprietary data. This announcement was made at Nvidia GTC, Nvidia’s annual technology conference, which this year prominently features discussions on AI and agentic models tailored for enterprise applications.

This strategic move is particularly significant for Mistral, a company that has consistently cultivated its business around corporate clientele, even as competitors like OpenAI and Anthropic have achieved substantial consumer market penetration. According to CEO Arthur Mensch, Mistral's unwavering dedication to the enterprise sector is proving highly effective, with the company projected to exceed $1 billion in annual recurring revenue this year.

A core tenet of Mistral's deepened commitment to the enterprise market is to provide companies with enhanced autonomy over their data assets and AI infrastructure.

"What Forge does is it lets enterprises and governments customize AI models for their specific needs," stated Elisa Salamanca, Mistral’s head of product, in an interview with TechCrunch.

While numerous companies within the enterprise AI landscape assert similar offerings, the majority typically concentrate on fine-tuning pre-existing models or integrating proprietary data via methods such as retrieval augmented generation (RAG). These prevalent strategies do not involve a fundamental retraining of the models; rather, they adapt or query them dynamically at runtime using organizational data.

Mistral, conversely, asserts that its platform allows companies to train AI models entirely from the ground up. Theoretically, this capability could mitigate several limitations inherent in more common methodologies — for instance, facilitating superior processing of non-English or highly specialized domain data, and offering greater granular control over model behavior. Furthermore, it could empower organizations to develop agentic systems utilizing reinforcement learning, thereby diminishing dependence on external model providers and circumventing associated risks such as unforeseen model alterations or deprecations.

Forge clients gain the ability to construct their custom models leveraging Mistral’s extensive library of open-weight AI models, which encompasses smaller iterations like the recently launched Mistral Small 4. Timothée Lacroix, Mistral co-founder and chief technologist, highlights that Forge is designed to extract greater value from their existing model portfolio.

"The trade-offs that we make when we build smaller models is that they just cannot be as good on every topic as their larger counterparts, and so the ability to customize them lets us pick what we emphasize and what we drop," Lacroix explained.

While Mistral offers guidance on optimal models and infrastructure, ultimate decision-making authority rests with the customer, Lacroix confirmed. For teams requiring more than mere advice, Forge also provides access to Mistral’s dedicated team of forward-deployed engineers (FDEs). These experts embed directly with clients, assisting in identifying pertinent data and tailoring solutions to specific requirements — an operational model inspired by companies such as IBM and Palantir.

"As a product, Forge already comes with all the tooling and infrastructure so you can generate synthetic data pipelines," Salamanca noted. She further elaborated, "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."

Mistral has already extended access to Forge to a diverse group of partners, including Ericsson, the European Space Agency, Italian consulting firm Reply, and Singapore’s DSO and HTX. Among its early adopters is ASML, the prominent Dutch chipmaker, which notably spearheaded Mistral’s Series C funding round last September, valuing the company at €11.7 billion (approximately $13.8 billion at that time).

These strategic partnerships are indicative of the primary use cases Mistral envisions for Forge. According to Marjorie Janiewicz, Mistral’s chief revenue officer, these encompass governments requiring models tailored to their specific language and cultural nuances; financial institutions navigating stringent compliance mandates; manufacturers with distinct customization imperatives; and technology firms needing to fine-tune models to their unique codebases.