The failure of most enterprise AI initiatives is not attributed to a lack of technological capability, but rather to the inability of the models being utilized to comprehend their business context. Often, these models are trained on general internet data rather than the extensive internal documents, workflows, and knowledge accumulated over decades. 

This disconnect presents an opportunity for Mistral, a French AI startup. On Tuesday, the firm unveiled Mistral Forge, a platform enabling enterprises to develop tailored models trained specifically on their proprietary data. Mistral introduced the platform during Nvidia GTC, Nvidia’s yearly technology conference, which is heavily centered on AI and autonomous models for enterprises this year.

This is a strategic initiative for Mistral, a company that has focused its efforts on corporate clientele while competitors like OpenAI and Anthropic have gained significant ground in consumer adoption. CEO Arthur Mensch asserts that Mistral’s dedication to the enterprise sector is yielding results: The organization is poised to surpass $1 billion in annual recurring revenue this year.

A significant aspect of intensifying its commitment to enterprises is facilitating greater control over their data and AI systems, according to Mistral. 

“Forge allows enterprises and government bodies to personalize AI models according to their unique requirements,” stated Elisa Salamanca, Mistral’s head of product, in an interview with TechCrunch. 

Numerous companies within the enterprise AI domain already claim to provide comparable functionalities, but the majority concentrate on refining existing models or layering proprietary data through methods like retrieval augmented generation (RAG). Unlike these approaches that do not fundamentally retrain models, Mistral’s method enables adaptation or querying at runtime using organizational data.

In contrast, Mistral emphasizes that it empowers businesses to train models from the ground up. Conceptually, this approach could mitigate some limitations faced by more prevalent methods — for instance, enhancing the treatment of non-English or highly specialized data, along with affording greater control over model functionality. Additionally, it may enable organizations to train autonomous systems using reinforcement learning, thereby decreasing dependence on third-party model suppliers and mitigating risks related to model modifications or discontinuation. 

Customers of Forge can develop their tailored models utilizing Mistral’s extensive collection of open-weight AI models, which encompasses smaller models such as the newly launched Mistral Small 4. Timothée Lacroix, Mistral’s co-founder and chief technologist, indicated that Forge aims to unlock additional value from existing models. 

“The compromises made when developing smaller models imply they cannot excel in every domain as their larger counterparts do, allowing customization to focus on specific strengths and weaknesses,” Lacroix articulated. 

While Mistral offers counsel on model and infrastructure selection, the ultimate decisions remain with the client, Lacroix noted. Furthermore, for teams seeking more than mere advice, Forge is supported by Mistral’s team of embedded engineers who collaborate directly with clients to identify suitable data and adapt to their specific needs — a strategy inspired by companies like IBM and Palantir. 

“Forge is a complete product, equipped with all necessary tools and infrastructure to create synthetic data pipelines,” Salamanca mentioned. “However, determining how to create effective evaluations and ensuring an adequate quantity of data is an area where enterprises often lack expertise, which is what the forward-deployed engineers provide.” 

Mistral has already made Forge accessible to partners such as Ericsson, the European Space Agency, the Italian consulting firm Reply, and Singapore’s DSO and HTX. Early adopters comprise ASML, the Dutch semiconductor manufacturer that previously led Mistral’s Series C funding round at a valuation of €11.7 billion (around $13.8 billion at that time).

These collaborations showcase what Mistral anticipates to be the primary applications for Forge. As per Marjorie Janiewicz, Mistral’s chief revenue officer, these applications include governmental entities needing to customize models for their specific language and cultural contexts; financial institutions with strict compliance needs; manufacturers requiring personalized solutions; and technology firms that aim to fine-tune models based on their codebases.