Unique Research · Brief · December 9, 2024
Meta announced Llama 3.3 70B — text-only — claiming the performance of its largest Llama, Llama 3.1 405B, at lower cost. Generative AI VP Ahmad Al-Dahle on X: post-training advances deliver “significantly better core performance at lower cost.”
A chart showed Llama 3.3 70B beating Google Gemini 1.5 Pro, OpenAI GPT-4o, and Amazon’s new Nova Pro on benchmarks including MMLU. A Meta spokesperson said to expect gains in math, common sense, instruction following, and tool use.
Downloadable from Hugging Face and Llama’s site — latest move in Meta’s “open” model strategy for multi-use and commercial apps. Terms restrict platforms with >700M MAU (special license required). Strict “openness” debates aside, Meta claims 650M+ Llama downloads. Internally, Zuckerberg says Meta AI runs entirely on Llama and nears 600M MAU — aiming to be the world’s most-used AI assistant.
Openness cuts both ways: reports last November that Chinese military researchers used Llama for a defense chatbot; Meta later offered Llama to U.S. defense contractors. Meta also frets about EU AI Act predictability for open releases, and GDPR: training on public Instagram/Facebook posts from users who did not opt out. EU regulators earlier this year asked Meta to pause European-user training while assessing GDPR; Meta complied while backing an open letter for a “modern reading” of GDPR that does not “reject progress.”
Like peers, Meta is scaling compute: Wednesday it announced a $10B AI data center in Louisiana — its largest. On the Q4 call (August), Zuckerberg said Llama 4 will need ~10× the compute of Llama 3; Meta has already procured a cluster of >100k NVIDIA GPUs. Capex rose nearly 33% YoY in Q2 2024 to $8.5B (from $6.4B), driven by servers, data centers, and networking.
— END —