On Thursday, Google unveiled the complete version of Gemma 3n, the latest addition to its Gemma 3 series of artificial intelligence models. Initially introduced in May, this new iteration has been tailored for on-device applications and incorporates significant architectural enhancements. Remarkably, the large language model (LLM) can operate locally with just 2GB of RAM, making it feasible for deployment on smartphones equipped with AI-capable processors.
Gemma 3n as a Multimodal AI Model
In a blog post, the tech giant from Mountain View shared details about the full release of Gemma 3n. This model follows the introductions of the Gemma 3 and GemmaSign models, and it is part of the growing Gemmaverse. As an open-source project, Google has made the model weights and associated cookbooks available to the community. Additionally, it operates under a permissive Gemma license, facilitating both academic and commercial uses.
Gemma 3n functions as a multimodal AI model, inherently supporting various input types, including image, audio, video, and text. However, it is limited to producing text outputs. The model boasts multilingual capabilities, accommodating 140 languages for text and 35 languages for inputs that involve multiple modalities.
According to Google, Gemma 3n features a “mobile-first architecture” based on the Matryoshka Transformer, or MatFormer, architecture. Described as a nested transformer reminiscent of Russian nesting dolls, this design allows for innovative training methods with AI models of varying parameter sizes.
The model comes in two variants: E2B and E4B, which refer to effective parameters. Despite comprising five billion and eight billion total parameters, the effective parameters are limited to two and four billion, respectively.
This efficiency is achieved through a method called Per-Layer Embeddings (PLE), which only necessitates loading essential parameters into fast memory (VRAM), while the remainder is managed through secondary embeddings accessed by the CPU.
Using the MatFormer architecture, the E4B model integrates the E2B framework, allowing for simultaneous training of both models. This provides users with the option to utilize E4B for complex tasks or E2B for quicker outputs, all without perceivable differences in processing quality or output.
Moreover, Google is enabling users to develop custom-sized models by adjusting specific internal components. To facilitate this customization, the company is rolling out the MatFormer Lab tool, which will assist developers in experimenting with various combinations to identify ideal model sizes.
Currently, Gemma 3n can be downloaded through Google’s Hugging Face listing and Kaggle listing. Users can also explore Gemma 3n via Google AI Studio. Notably, deployments of Gemma models can be executed directly to Cloud Run from AI Studio.
