Gemma 3, Google's open-source AI model, enhances vision-language understanding and efficiency with features like a new Sigmoid loss for pre-training and 'Pan & Scan' algorithms for imagery. It supports 4B, 12B, and 27B parameter models and utilizes bi-directional attention for a deeper contextual grasp. Key improvements include KV-cache memory reduction and a compact token representation for images, which optimize resource use while processing visuals. These advancements position Gemma 3 as a notable upgrade over its predecessor in managing long contexts effectively.
Gemma 3 introduces advanced vision-language understanding features, improved multi-linguality, and efficient memory management, enhancing performance on visual and text-based tasks.
The new model implemented a custom Sigmoid loss for vision-language tasks and utilizes a 'Pan & Scan' algorithm for better handling of high-resolution images.
This model's bidirectional attention allows for a deeper understanding of tasks where the entire text is present, significantly enhancing context comprehension.
Architectural changes effectively reduce KV-cache memory overhead during inference, making Gemma 3 more efficient than its predecessor while handling longer contexts.
Collection
[
|
...
]