Mistral AI: A European Contender in Generative AI
Mistral AI, a French start-up founded by former researchers from Meta and Google DeepMind has just made two key announcements. First, the company has closed its Series A funding round,
securing €385 mm (~$415 mm) less than six months after its debut and €105 mm seed funding announcement. The funding, led by Andreessen Horowitz and Lightspeed, brings Mistral AI’s valuation to nearly $2 billion.
Second, as is becoming common with these splashy funding announcements, the company is releasing Mixtral 8x7B, a high quality model that uses a “sparse mixture of experts” approach to achieve performance comparable to much larger models (like GPT-3.5 and Llama 2 70B) while being significantly faster (6x faster inference) and computationally efficient. Beta access to Mixtral 8x7B and other models by the company is available through three generative endpoints on the Mistral platform.
Mistral AI’s founding manifesto lays out a vision for a Europe-centric, open-source development of LLM technology. In September, they released “Mistral 7B,” under an Apache 2.0 license. This smaller model, with 7 billion parameters, can be run locally on most modern machines (I run it using Ollama on a 2017 iMac) and offers performance close to Meta’s Llama 2 13b (which, by the way, can also be run with Ollama on my iMac.)
Today’s launch of API access marks an important step towards commercialization for Mistral AI. The platform provides developers with access to Mistral AI’s LLMs through paid APIs (similar to OpenAI), allowing them to integrate the models into their own applications and services. This approach allows Mistral AI to monetize its technology while still ensuring its models remain accessible to a wider audience.
The platform offers three models: Mistral-tiny (based on the Mistral 7B model), Mistral-small (the new Mixtral 8x7B model with enhanced efficiency), and Mistral-medium (the highest-performing model currently available only through the API).
Beyond its technological achievements, Mistral AI has also been an active advocate for responsible AI development. The company claims that open-source technology fosters transparency and collaboration, leading to safer and more robust AI. Mistral AI played a significant role in shaping the EU’s AI Act, advocating for total exemption for foundational models and focusing regulation on use cases and products. The recently agreed-upon AI Act reflects this approach, requiring transparency from developers while avoiding unnecessary restrictions on foundational models themselves. Contrasting US-tech’s poor track record on consumer privacy and data handling with the EU’s leadership in protecting consumer data, I think this is a positive development, albeit a tiny first step.
I believe Mistral AI is positioned to become a major force in generative AI – keep an eye on them.