Meta Platforms Retakes Its Open-Source Lead

Meta Platforms made a splash Thursday with the release of Llama 3, the latest version of its popular open-source large language model, though we stole some of the thunder with our scoop earlier this month. Developers have a lot to say about it—most of it good.

For the most part, the 8-billion and 70-billion parameter versions of Llama 3 outperform similarly sized LLMs from Google, Mistral and Anthropic, according to the company’s blog post. Independent reviewers are also impressed. The 70-billion parameter version of Llama 3 is the best-performing open-source LLM available, according to LMSYS’ leaderboard, which says only frontier closed-source models from OpenAI, Anthropic and Google perform better. (As a reminder, parameters are the “settings” that determine how AI models respond to questions.)

Plus, Meta says that the 405-billion parameter version of Llama 3, which it is still training, is nearing the performance of those closed-sourced models. As of last Monday, the 405-billion parameter version of Llama 3 scored just behind Google’s Gemini Ultra, Anthropic’s Claude 3 Opus and GPT-4 on the MMLU benchmark, according to Meta’s blog post and Papers With Code’s leaderboard.