Meta releases Llama 3 and puts its assistant everywhere

Meta released the first Llama 3 models: 8-billion and 70-billion parameter versions, both pretrained and instruction-tuned, available for download under Meta’s community licence. Meta said the models set a new state of the art for openly available LLMs at those sizes and, in human evaluations across a set of prompts covering coding, reasoning and creative-writing use cases, that the 70B model was preferred over Claude Sonnet, Mistral Medium and GPT-3.5. The company also disclosed that a substantially larger model — over 400 billion parameters — was still training, and previewed added multilingual and multimodal capability and a longer context window for that later release.

The same announcement doubled as a product rollout. Meta AI, built on Llama 3, began appearing as a built-in assistant across Facebook, Instagram, WhatsApp, Messenger and the web, reaching a user base numbering in the billions by default rather than by opt-in download — a distribution advantage no other open-weight model family could match, since none of Llama’s competitors sat inside a social-media platform with that reach. Meta also previewed multimodal Meta AI features for its Ray-Ban smart glasses.

Llama 3 continued Meta’s strategy, begun with Llama 2 in 2023, of releasing competitive open-weight models for free while building a first-party consumer product on top of them — a combination that pressured both closed labs, who had to justify keeping weights private against models of comparable quality available for download, and the open-source ecosystem, which increasingly organised itself around fine-tuning and serving Llama rather than competing frontier open releases from smaller labs. The 400B-plus model referenced in this post shipped later in 2024 as Llama 3.1 405B, at the time the largest openly available model from a major lab.