Meta Llama 3.3 70B delivers Llama 3.1 405B performance in a smaller, cost-effective package

Meta continues to bet on its 'open release' strategy with Llama 3.3 70B, a text-only (input and output) model with a performance approaching that of Meta's larger model, Llama 3.1 405B, introduced earlier this year as an "open" alternative to Anthropic's Sonnet 3.5 or OpenAI's GPT-4 and 4o. Of course, both Anthropic and OpenAI have kept improving their offerings, which may make the comparison slightly outdated. However, achieving a performance that counted as state-of-the-art less than half a year ago is a feat that deserves recognition, even more so if offered under a permissive license.

Ahmad Al-Dahle, VP of generative AI at Meta, posted on X that Llama 3.3 70B would make competitive performance even more accessible to the open-source community, given its size and significantly lower cost. Al-Dahle's post included the following chart, which not only shows Llama 3.3 70B outperforms GPT-4o, Llama 3.1 405B, and even the recently launched Amazon Nova Pro in some popular industry benchmarks, but it reveals that Llama 3.3 70B costs only $0.01 per million input tokens—against $1.0 for Llama 3.1 405B— and $0.4 per million output tokens—against $1.8 for Llama 3.1 405B.

Mark Zuckerberg took to Instagram to share the news about Llama 3.3 70B, confirming this model is the last Llama 3 release, as Meta prepares to launch Llama 4 in 2025. Llama models have become wildly popular: Meta estimates Llama models have been downloaded an estimated 650 million times, and according to Zuckerberg, they are also on the way towards becoming an industry standard. Zuckerberg also stressed that Meta AI, Meta's AI-powered assistant, is on track to become the most used AI assistant in the world with 600 million monthly active users. Finally, the announcement closed with an acknowledgment of Meta's upcoming Louisiana data center, which will be used to train future versions of Llama.