Gemma 3n

What it is

Gemma 3n is a mobile-first open model in the Gemma family, released in full on 26 June 2025.^[1] Google's model page describes it as a generative AI model optimised for everyday devices such as phones, laptops and tablets.^[2] Almost everything else about it follows from that setting. A model that has to run on a phone works inside a fixed memory budget and limited compute, and the design choices below are answers to those limits.

Two sizes, two memory footprints

Gemma 3n comes in two sizes, E2B and E4B. They hold 5 billion and 8 billion raw parameters respectively, but run in about 2 GB and 3 GB of memory.^[1] The gap between the two kinds of number is deliberate. Google's Gemma 3n model page says that executing the E2B model loads over 5 billion parameters, while the effective memory load is just under 2 billion, at 1.91B.^[2] For a device, the second figure is the one that decides whether the model fits at all.

MatFormer and Per-Layer Embeddings

Two mechanisms produce that gap. The first is MatFormer, short for "Matryoshka Transformer": an architecture that nests a fully functional smaller model inside a larger one.^[1] The name comes from the Russian nesting dolls, and the idea is the same — a model that contains a smaller model, which can be used on its own. The second is Per-Layer Embeddings, which let much of the parameter count be loaded on the CPU.^[1] Together they explain how the raw parameter counts and the running memory figures come apart: much of what the model carries can be loaded where it is needed rather than held in the same place all at once.

What it can take in

Gemma 3n natively accepts image, audio, video and text input and produces text output.^[1] Text support covers 140 languages, and multimodal understanding covers 35 languages.^[1] The output side is text only, which is worth stating plainly: the model reads pictures, sound and video, and answers in writing.

Benchmark

Google said the E4B model was the first model under 10 billion parameters to score above 1300 on LMArena.^[1] That is a statement about relative standing among small models, made by the model's publisher.

How to run it

Google said Gemma 3n offers multiple deployment options, including the Google GenAI API and Vertex AI.^[1] The developer guide listed runtime and serving tools including Ollama, MLX, llama.cpp, vLLM and SGLang.^[1] That spread matters for a model pitched at devices: the same weights run under several different software stacks rather than tying users to one.

Where it sits

Within the Gemma family, Gemma 3n is the member built around the device, and its specifications are written in terms of the memory a phone has to spare. For anyone following small models, what is interesting is the trade-off it makes explicit: a model's size can be reported as a parameter count or as the memory it needs on the device in front of you, and Gemma 3n is designed so that those two figures differ by a wide margin.