Key Takeaways
The rise of large language models (LLMs), especially OpenAI’s GPT range, contributed to the end of a golden age of open-source AI, Mistral CEO Arthur Mensch claimed in a recent episode of the a16z Podcast .
However, he pointed to the emergence of smaller, open models like NeMo, which Mistral developed in collaboration with Nvidia, as evidence of a renaissance in open-source AI.
Speaking to A16z’s Anjney Midha and Nvidia CEO Jensen Huang, observed that “between 2010 and 2020” the world’s leading AI labs were mostly “building on top of each other” in an open-source development paradigm.
Yet, “that’s something that kind of disappeared with the first large language models, from OpenAI in particular,” he added.
From GPT-3 onward, OpenAI has kept the code and weights of its AI models private.
Its peers in the space, including Google and Anthropic, have followed suit, and today, the largest models on the market are strictly proprietary.
While OpenAI’s largest models now reach over a trillion parameters, smaller models like Mistral’s are increasingly competitive.
Unlike the Big Tech giants, small and mid-sized model developers embrace open-source.
Mensch said that bringing back the open-source philosophy that characterized AI development in the 2010s is “the reason why we created Mistral.”
Alongside Mistral, companies such as Meta, Alibaba, and Deepseek continue to contribute to open-source AI development.
As these initiatives advance, the technology evolves iteratively. Deepseek built its R1 model on top of Alababa’s Qwen, which in turn borrowed from Meta’s Llama.
Asdeep Mensch observed, “everybody benefits from it.”