It’s been just one week since we put Meta Llama 3 in the hands of the developer community, and the response so far has been awesome. With the release of our initial Llama 3 models, we wanted to kickstart the next wave of innovation in AI across the stack—from applications to developer tools to evals to inference optimizations and more, and we’re already seeing amazing things. From projects that have doubled Llama 3’s context window, to work on quantization, web navigation, tool use, low precision fine tuning, local deployment, and so much more, the reception from the community has been tremendous.

Since our launch last Thursday, the models have been downloaded over 1.2 million times, with developers sharing over 600 derivative models on Hugging Face. The Llama 3 GitHub repo has already passed 17,000 stars, and Llama 3 70B Instruct is tied for first for English-only evals on the LMSYS Chatbot Arena Leaderboard, and sits at six overall making it the highest ranked openly available model, just behind closed proprietary models.

Our partners across silicon, hardware, and cloud providers have begun to deploy Llama 3 to their users, and we’ve seen the community quickly fine-tune Llama 3 for their individual needs. For example, at Yale’s School of Medicine, teams alongside the EPFL School of Computer and Communication Sciences fine-tuned Meta Llama 3 within 24 hours of release, introducing Llama-3[8B]-MeditronV1.0—the first fine-tuned Llama 3 8B for medicine. This new model outperforms all state-of-the-art open models within its parameter class on standard benchmarks such as MedQA and MedMCQA. You can read more about how Yale and the EPFL built their first version of Meditron on top of Llama 2 here.

As we shared at launch, this is just the beginning for Llama 3. Over the coming months, we’ll release models with new capabilities including multimodality, the ability to converse in multiple languages, a longer context window, and enhanced overall capabilities. We’re excited to share more soon and see the next-wave of community-driven innovation.