Sep 26, 2024 10:45:00

Meta releases 'Llama 3.2', with improved image recognition performance and a smaller version for smartphones

Meta announced the AI model ' Llama 3.2 ' on September 26, 2024. In addition to the medium-sized models 'Llama 3.2 90B' and 'Llama 3.2 11B' with improved image recognition performance, Llama 3.2 also includes the small-sized models 'Llama 3.2 3B' and 'Llama 3.2 1B' optimized for local operation on smartphones.

Llama 3.2

https://www.llama.com/

Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

◆“Llama 3.2 90B” & “Llama 3.2 11B”
Llama 3.2 90B and Llama 3.2 11B have been trained to improve image recognition performance while maintaining the text-related features of Llama 3.1, allowing for high accuracy inference results when inputting images and text simultaneously.

Below is a table comparing the benchmark results of 'Llama 3.2 90B' and 'Llama 3.2 11B' with 'Claude 3 Haiku' and 'GPT4o-mini'. 'Llama 3.2 90B' outperforms rival models in many tests, and 'Llama 3.2 11B' also records a high score for a relatively small model.

◆“Llama 3.2 3B” & “Llama 3.2 1B”
'Llama 3.2 3B' and 'Llama 3.2 1B' are lightweight models designed for operation on smartphones, and can perform operations such as 'summarization,' 'rewriting,' and 'sentence generation' on low-power smartphone chips. In addition, when developing 'Llama 3.2 3B' and 'Llama 3.2 1B,' it was said that they worked closely with Arm, Qualcomm, and Mediatek.

In tests of rewriting text on a smartphone, the Llama 3.2 3B and Llama 3.2 1B outperformed lightweight models such as the Phi 3.5 3.8B and Gemma 2 2.6B.

Below is a table listing the benchmark results for 'Llama 3.2 3B', 'Llama 3.2 1B', 'Gemma 2 2B IT', and 'Phi-3.5-mini IT'. 'Llama 3.2 3B' outperforms its rivals in many tests, and 'Llama 3.2 1B' also outperforms its rivals in some tests.

Various models of Llama 3.2 can be downloaded from the following links. However, only the quantized models of 'Llama 3.2 3B' and 'Llama 3.2 1B' are available, and details of the unquantized models will be released soon.

Meta-llama (Meta Llama)
https://huggingface.co/meta-llama