Publicly released benchmark scores

Meta Unveils Llama 3.2 AI Models, 90B Outperforms GPT-4o-mini

Sep 25, 2024, 05:46 PM

Meta has unveiled Llama 3.2, a new family of AI models featuring both lightweight and multimodal capabilities. The new models come in various sizes, including 1B, 3B, 11B, and 90B parameters. The 1B and 3B models are optimized for edge devices, while the 11B and 90B models are multimodal, capable of processing both text and images. Llama 3.2 models outperform many existing models on industry benchmarks, including GPT-4o-mini and Claude Haiku. The models are open-source, providing developers with access to powerful AI tools that rival closed-source alternatives. The models have been trained on up to 9 trillion tokens and support a 128K context length. The 1B model scored 49.3 on the MMLU benchmark, while the 3B model scored 63.4. The release was announced at Meta Connect 2024 and is expected to have significant implications for both edge computing and vision-based AI applications.