Although it can sometimes seem as if Large Language Model (LLM) development is a two horse race between OpenAI and Google, a burgeoning field of Artificial Intelligence (AI) startups has become increasingly competitive. Away from the Big Tech giants, developers like Mistral form an alternative AI wave, returning it to its open-source roots.
A team of ex-Meta and Google researchers founded Mistral earlier this year. On December 10, it’s $415 million Series A funding round threw the French startup into the limelight. At a $2 billion valuation, the latest investment round positions Mistral as one of the giants of European AI and sets it up to take on some of the largest technology companies in the world.
Since its inception in May, Mistral’s founders have planted the company squarely in the tradition of free and open-source software.
With a commitment to “open science, community and free software,” the firm releases its AI models and deployment tools under permissive licenses, granting a degree of access and affordability that other players in the space have increasingly abandoned in recent years.
As Mistral founder Arthur Mensch explained in a recent podcast , modern machine learning (ML) technology is built on open-source foundations and for many years the free flow of information was the engine of AI innovation.
However, “all of a sudden in 2020 with GPT-3, this tide reversed”, referring to OpenAI’s closed approach to distributing LLM services
For the likes of Google and OpenAI, charging for the use of their latest AI platforms has proven extremely lucrative. To protect their commercial interests, both companies strictly control access to the underlying models, requiring developers to interact with them via Application Programming Interfaces (API) instead.
In contrast, people can download open AI models, like Mistral’s, for free.
To date, the company has put off monetization to focus on building a platform developers want to use. But the firm probably wouldn’t have been able to raise hundreds of millions of dollars if investors like Andreessen Horowitz (a16z) didn’t believe it was commercially viable.
As Mensch noted in conversation with a16z General Partner Anjney Midha: “Many businesses […] have a very strong open source community and also a very good offer of services and that’s what we want to be.”
In a September blog post , the Mistral team expanded on their proposed business model:
It said: “We’re committing to release the strongest open models in parallel to developing our commercial offering.”
Taking its cue from open-source pioneers like Red Hat, Mistral intends to develop optimized, proprietary versions of its models to distribute as paid-for solutions.
So far, the open-source AI community has responded positively to Mistral’s AI models. They have been lauded by developers for their efficiency and lower cost of use.
But in a previous era, OpenAI was also committed to making its products freely available. Yet since transitioning to a for-profit business model in 2019, the firm has abandoned its original mission.
Likewise, Google has historically been one of the largest contributors to open-source projects. The company’s commitment to open innovation has always had its limits, and its most advanced AI models have become jealously guarded secrets.
Ultimately, AI has started to reshape notions of openness, and business models are still adapting to the new technology
The freedom to run, study, modify and redistribute code is central to the open-source movement. But contemporary neural networks lack the kind of human-readable codebase free software licenses were designed for.
Developers like Mistral are exploring new licensing models based on the idea of “Open Weights.” The concept remains underdeveloped, however, and the future of it isn’t yet clear where the future of open-source AI lies.