Posted December 10, 2023

AI should be open.

Most of the core systems powering modern computing are now open source: server operating systems (Linux), virtualization (KVM, Docker), databases (Postgres), data analytics (Spark, DBT), programming languages (Javascript, Python, C), web servers (NGINX) — the list goes on.

Why? The community.

No single engineering team can anticipate every user’s needs or catch every bug. But by giving everyone access to a piece of software — and encouraging anyone to contribute back code — open source projects build a flywheel to improve features, stability, performance, security, and more. Communities build cheaper, faster, and safer software infrastructure together. AI is no different.

That’s why we’re thrilled to announce our Series A investment in Mistral.

Mistral is at the center of a small but passionate developer community growing up around open source AI. These developers generally don’t train new models from scratch, but they can do just about everything else: run, test, benchmark, fine tune, quantize, optimize, red team, and otherwise improve the top open-source LLMs. Community fine-tuned models now routinely dominate open source leaderboards (and even beat closed source models on some tasks).

We think this is the most promising path to achieve robust, widely adopted, and trusted AI systems, and that Mistral is the leading independent team on this path.

In the seven months since the company was founded, Mistral has released the most powerful LLMs in both major “weight classes” (mistral-7b and mixtral). These models are also dramatically more efficient than most alternatives, have less restrictive filtering, and demonstrate rapid improvement in the underlying architectures (e.g., sparse mixture-of-experts). As a result, most model fine-tuners and infrastructure projects are now designing their roadmaps around Mistral models.

This is an incredible accomplishment for a young company. But it’s also not surprising. Led by Arthur Mensch, Guillaume Lample, and Timothee Lacroix, Mistral is a brilliant team of former Meta and Deepmind researchers and engineers, who worked on a number of the top open source and openly published models. Notably, they were core contributors to the other highly performant open source model family (Llama), and to the project that pioneered a focus on small, data-rich models (Chinchilla).

Mistral is dedicated to the mission of bringing open source AI models to the world. We believe they are picking up where other AI labs and big companies have left off, fostering open research, open models, and open collaboration in the AI community.

You can learn more about their plans and latest models in this episode of the a16z Podcast.