Mistral launches Forge to help enterprises build their own AI models

Mistral has introduced Forge, a new platform aimed at helping enterprises move beyond generic AI systems by enabling them to train and adapt models on proprietary data.

Today’s AI systems are largely developed using open internet data and are built to handle a wide variety of general tasks. However, enterprises depend on deeply embedded internal knowledge, including proprietary processes, regulatory requirements, custom software environments, and the accumulated experience of their organizations.

“Forge bridges the gap between generic AI and enterprise-specific needs,” Mistral said in a statement. “Instead of relying on broad, public data, organizations can train models that understand their internal context embedded within systems, workflows, and policies, aligning AI with their unique operations.”

Forge supports multiple stages of the model lifecycle, Mistral said, including pre-training on internal datasets, post-training for specific tasks, and reinforcement learning to align models with internal policies and operational requirements.

The company said Forge is already being used by organizations including ASML, Ericsson, and the European Space Agency.

Mistral is also emphasizing data control, saying enterprises can retain ownership of both models and underlying data, an issue for companies concerned about reliance on third-party AI providers.

The move puts Mistral in closer competition with players like OpenAI and Anthropic, which have so far focused more on general-purpose models and enterprise integrations.

Enterprise feasibility and adoption

While the idea may be appealing, not everyone is convinced it will see widespread adoption. Concerns around cost and readiness could limit custom model development to niche use cases.

Building models from scratch will remain realistic only for a “small set of large enterprises with strong AI talent, deep budgets, and specific data advantages,” according to Tulika Sheel, senior vice president at Kadence International.

“For most organizations, fine-tuning and RAG [Retrieval Augmented Generation] will continue to be more practical and cost-efficient,” Sheel said. “Where Mistral’s approach becomes relevant is in highly regulated or domain-specific sectors that need full control over data, models, and outputs.”

Sheel added that fully customized models are most relevant in compliance-heavy industries, multilingual environments, and highly specialized workflows such as legal, healthcare, and financial analysis. In these scenarios, generic models augmented with retrieval techniques may fall short in delivering the required nuance and consistency.

Others said Mistral’s offering will be useful only when enterprises have a clear strategic understanding of how they want to deploy AI.

“I think we are not there yet, with enterprises still figuring out AI,” said Faisal Kawoosa, founder and chief analyst at Techarc. “It’s good that they have introduced this concept, and my sense is enterprises will experiment with it for now. But I don’t see any serious deployments for at least the next two years, by which time enterprises may have greater clarity on AI in their businesses.”

Still, analysts suggest there is scope for such offerings, as data sovereignty is becoming increasingly important, particularly in regions such as Europe and the Middle East, and in sectors like finance, legal, quantum computing, and healthcare.

“The frontier models fine-tuned for these sectors, whether based on open or proprietary foundation models, do not offer the desired level of sovereignty, and Mistral is trying to address that problem with Forge,” said Neil Shah, VP for research at Counterpoint Research. “Fully customized, pruned, and optimized models can deliver more accurate and relevant outputs compared to the RAG approach currently used in frontier models.”