one year on
Meta releases Llama 3.1 405B, claims first frontier-level open source AI model
Mark Zuckerberg argues open source AI will become the industry standard, releasing a 405-billion-parameter model that he says has roughly 50% the inference cost of closed models like GPT-4o.
Meta today released Llama 3.1 405B, which CEO Mark Zuckerberg calls ‘the first frontier-level open source AI model,’ alongside updated 70B and 8B variants. In a lengthy blog post, Zuckerberg argues that open source AI will follow the trajectory of Linux, eventually surpassing closed models in performance, security, and ecosystem breadth.
The 405B model is available on major cloud platforms including AWS, Azure, and Google Cloud, with inference costs claimed to be roughly 50% of using GPT-4o. Zuckerberg says the open approach lets organizations fine-tune and distill models without vendor lock-in or exposing sensitive data. He also defends openness on safety grounds, arguing transparency enables better scrutiny.
The record
One year later — open only if you can handle spoilers
By mid-2026, Llama 3.1 405B became a widely used baseline for fine-tuning and distillation, though the claim of being 'frontier-level' was contested by proprietary model makers. The open-source AI ecosystem continued to grow, with Meta releasing several subsequent Llama iterations, while closed models like GPT-4o remained dominant in some benchmarks.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy