From the source
Gemma 3n, a natively multimodal on-device model supporting image, text, audio, and video inputs, is now fully available in the open-source ecosystem.
Two sizes (E2B and E4B) are released with base and instruct variants, achieving competitive performance (e.g., E4B scores 1300+ on LMArena) and efficient memory usage (E2B runs in ~2GB GPU RAM, E4B in ~3GB).
The models are supported by transformers, timm, MLX, llama.cpp, and other libraries.
