Summary Bullets:
• Startup Mistral AI was founded nine months ago in April 2023 and was the talk of Davos amid speculation about the future of AI.
• The French company recently released its most powerful model yet, Mixtral 8x7B, under the Apache 2.0 license.
The two-horse race between Microsoft-backed OpenAI and Google for the GenAI crown may be a thing of the past. French startup Mistral AI seems poised to disrupt the status quo and was the center of generative AI talks at the World Economic Forum in Davos (Switzerland) last week.
The company, based in Paris (France), was created a mere nine months ago, and it reached an EUR2 billion valuation in December 2023 after a founding round of close to EUR400 million. Following a familiar formula, it was founded by former employees of Google DeepMind’s and Meta’s research labs. CEO Arthur Mensch created advanced models at Google DeepMind, and co-founders Guillaume Lample and Timothée Lacroix honed their skills developing early iterations of Llama while at Meta.
The company has become a mighty contender in the competitive landscape of large language models (LLMs) made by US behemoths such as Google and Amazon. Of course, it is not the only one trying to do this; startups such as Anthropic, Cohere, and Hugging Face have been attempting just that with varying degrees of success. However, Mistral AI seems confident it can compete at the highest level and boasts superstar chipmaker Nvidia as an investor and partner, among other big names including Salesforce. Although not an investor, the company has also been singled out by Microsoft’s Satya Nadella, who says it is one of the top innovators building on the Microsoft Azure platform today.
Mistral AI has introduced several innovations into the traditional neural network ‘Transformer’ architecture, resulting in optimized memory, performance, response quality, and inferencing techniques. These novel approaches have given its open-source platforms a distinct edge. The company is not profitable yet, but it is already working on proof of concepts with several global corporations. Its open-source approach attracts a specific type of organization in highly regulated industries where firms are trying to experiment with GenAI while avoiding proprietary software for compliance reasons.
In September 2023, the company released its first model, Mistral 7B, under the open-source Apache 2.0 license, and made it available as a free download rather than via proprietary APIs. Despite its relatively small size of seven billion parameters, it punches above its weight in terms of performance and has a good overall value when it comes to operating costs. However, the company’s latest and greatest model, Mixtral 8x7B, is only accessible via an API; although it retains a permissive license under Apache 2.0. Mixtral 8x7B is out to compete with LLMs such as Llama 2. It utilizes an innovative technology dubbed ‘sparse mixture of experts’ that dramatically reduces inference times.
Mistral’s innovative acumen also represents European ambitions to nurture local talent at this historical moment for AI technologies in a continent with formidable education and research but sometimes weaker commercial acumen. The startup was involved in talks surrounding the preparation of the imminent AI Act, which is including a special clause to exempt foundation models from certain general rules to prevent stifling innovation. In the meantime, disrupting the hegemony of OpenAI looks like a mammoth task while the company quietly builds its GPT 5 frontier model. 2024 will sure be an exciting year for GenAI, and Mistral AI will be an important part of it, putting open-source models at the forefront of considerations about this technology.