How Mistral is driving growth through open source and enterprise AI

Mistral AI, the Paris-based generative artificial intelligence (GenAI) startup, is counting on open source and enterprise AI adoption to grow its business, its CEO and co-founder Arthur Mensch told attendees at the recent ATxSummit in Singapore.

During a fireside chat with Lew Chuen Hong, CEO of Singapore’s Infocomm Media Development Authority, Mensch said Mistral AI, founded in April 2023 by former Meta and Google researchers, aims to bring AI to enterprises and governments, enabling them to customise the technology and make it their own so they can control it without external dependencies.

“The natural way of doing that was to release open source models,” Mensch said, referring to Mistral AI’s first model released just four months after the company’s inception. “That was quite successful because it was the first model that could run on a laptop with decent performance. And since then, we’ve been committed to our open-source strategy of releasing stronger models.”

Mensch said open source has been a “huge advantage for business”, showing that powerful AI could be deployed on an organisation’s own hardware and private cloud along with control over its data. “The idea of being able to deploy locally and wherever you want has changed the way we look at the technology,” he added.

As for how Mistral AI balances its open source ethos with monetisation, Mensch acknowledged it’s a trade-off because the company wants to continue providing models relevant to the open source community, drive innovation and enable others to build on its research. “That’s the reason why we’re committed to open source and will remain so,” he said.

That said, Mistral AI monetises its work in different ways, such as offering public cloud services with application programming interfaces (APIs) for customers to build AI agents along with connectors to data sources, a platform deployable in an air-gapped environment, and full-scale products such as Le Chat, an AI assistant for work and personal use.

However, most of the company’s business comes from enterprise engagements, where it helps businesses to deploy AI applications. “We work with manufacturing, logistics, biotech and financial services companies to identify the most important use cases and do the integration work to deliver value very quickly,” Mench said.

A core tenet for Mistral AI is model efficiency without sacrificing performance. “The insight we had was that if you spent more on computing to compress more knowledge, you could actually make smaller models,” Mensch said, noting that “size matters because the larger the model, the more latency you will have”.

“When you’re building things with LLMs [large language models], you need them to be fast,” he added. “The faster they are, the more interesting things you can do. You can make them reason, do complex tasks in the background and still have decent latency.”

This efficiency extends to local deployments, with Mensch noting a trend towards hybrid systems “where you can have some compute happening at the edge for simple tasks and offload larger tasks to the cloud”. He also noted the increasing power of laptops and the capability of smaller models, such as the 24 billion parameter models, to act as effective agents locally for tasks such as coding.

For enterprises grappling with AI deployments, Mensch advised them to start with AI assistants for productivity, then moving to identify processes suited for automation. This involves designing custom AI systems designed to run complex processes using AI agents while involving people at the same time.

“Instead of people triggering agents to run, you have agents running at the process level and getting inputs from humans that are needed in the process loop. Once you’ve done that, you can progressively reallocate people to do things you still cannot do with AI,” he said.

Mistral AI recently released an agent API that allows users to connect tools, web search and code executors, with the company handling the orchestration.

“More of the orchestration will be moving to the server side and managed by us,” Mensch said. “We can manage the tokens for you, and you don’t need to worry about authentication and permissions which can be hard to figure out and maintain. We are progressively bringing these things onto our platform that can be self-deployed.”

On AI safety, particularly when it comes to AI agents, Mensch suggested sandboxing executed code and treating all external inputs as unsafe, pointing out the need for moderation and evaluation to ensure AI systems work as intended.

“When you’re calling a model, you don’t know what it’s going to output because there’s a random part to it that you need to manage,” Mensch said. “And we’re able to do it because we can make the entire system accurate enough so long as we can monitor and control the inputs.”

The company recently established an office in Singapore, signalling its growing ambitions in the Asia-Pacific region where governments and enterprises are warming to sovereign AI to minimise any dependency on technologies that might be cut off.

“We ship the software and make sure our customers and partners have access to it, so that even if we disappear, they still can maintain it,” Mensch said. “This aspect around sovereignty and strategic autonomy for such a core technology is super important in Europe. It’s turning out to be super important in Asia-Pacific as well, and that’s one of the reasons why we’re growing very fast here.”