Gemma 3n: on-device multimodal models with "effective" sizes
Gemma 3n arrived across open-source tools; its E2B and E4B models have 5B and 8B real parameters but need only 2B- and 4B-class memory, and accept image, audio, video and text.
What happened
Gemma 3n, previewed at Google I/O, became available across the main open-source tools: Transformers, timm, MLX, llama.cpp (text), transformers.js, Ollama and Google AI Edge.
The technical details
From the post:
- two sizes,
gemma-3n-E2Bandgemma-3n-E4B(base and instruct). E stands for effective:- real parameter counts are 5B and 8B
- thanks to memory-efficiency techniques, they need only about 2B / 4B of VRAM, and run in about 2GB and 3GB of GPU RAM
- natively multimodal: image, text, audio and video inputs
- besides the language decoder, an audio encoder and a MobileNet-V5 vision encoder
Why it matters
Parameter count and memory use are no longer the same thing. Models can now be designed specifically for phones and laptops, and the Gemma 4 posts later this year continue that line.
#llm#on-device#multimodal#google
Sources
- Gemma 3n fully available in the open-source ecosystem! — Hugging Face
This is a summary written from the linked public sources. Figures and numbers are as reported there.