Introducing Gemma 3n: The developer guide

Google DeepMind pulled back the curtain on Gemma 3n's full release this week, following last month's preview, and the pitch is straightforward: cloud-grade multimodal AI that actually fits on a phone. The E4B version, despite having 8 billion raw parameters, runs with a memory footprint closer to a 4B model thanks to a handful of architectural tricks. Its smaller sibling, E2B, needs as little as 2GB of memory. For context, that's the kind of headroom a mid-range Android device has to spare.

The cleverest piece here is something called MatFormer, short for Matryoshka Transformer. Google trained the 4B model so that a fully functional 2B model lives nested inside it, like a Russian doll. Developers can grab either size outright, or use a tool called MatFormer Lab to slice out custom sizes in between by adjusting feed-forward dimensions and skipping layers. It's a genuinely useful idea for anyone trying to squeeze a model onto specific hardware rather than accepting whatever size a lab decided to ship.

Two other engineering choices matter more than they sound. Per-Layer Embeddings shift a big chunk of the model's parameters onto the CPU instead of hogging GPU or TPU memory, so a model with 5 billion total parameters can run with only 2 billion actually sitting in the constrained accelerator RAM. And KV Cache Sharing speeds up how fast the model chews through long inputs, delivering a 2x improvement on prefill time versus Gemma 3 4B, which matters a lot if you're streaming audio or video and want a fast first response rather than an awkward pause.

On the multimodal side, Gemma 3n adds real speech-to-text and speech translation, built on a Universal Speech Model encoder that tokenizes audio every 160 milliseconds. Google says translation quality is particularly strong between English and Spanish, French, Italian and Portuguese. Vision comes from a new encoder called MobileNet-V5-300M, which processes up to 60 frames per second on a Pixel phone and, according to Google's own numbers, runs 13 times faster with quantization than the vision setup in Gemma 3, while using 46% fewer parameters.

Google is backing the release with a $150,000 Impact Challenge asking developers to build offline, on-device products with real-world use, and it's leaning on an already sprawling ecosystem, over 160 million downloads and support from Hugging Face, Ollama, llama.cpp, NVIDIA and others. Weights are up now on Hugging Face and Kaggle for anyone who wants to poke at it directly.