Google's Gemma 3n squeezes big multimodal AI into phones
Google just unveiled Gemma 3n, a mobile-first open model designed to run rich multimodal AI entirely on phones and laptops. Thanks to a new Per-Layer Embeddings technique, its 5B and 8B models operate with dynamic memory footprints of just 2GB and 3GB. Why it matters: it makes complex, private local AI actually fit on everyday hardware.
Gemma 3n is Google's new mobile-first open model, built to run real-time multimodal AI directly on phones, tablets, and laptops without pinging the cloud.
Released as an early preview, it comes in two parameter sizes: 5B and 8B. Instead of eating up your entire mobile RAM, both models are optimized to run entirely within the dynamic memory limits of everyday hardware.
What it can do
Gemma 3n lets you build local, completely private applications that can ingest audio, video, image, and text inputs simultaneously.
Because processing happens on-device, latency drops drastically and data never leaves the user's hands. It is particularly tuned for audio workloads: real-time speech transcription, on-device translation, and responsive voice-driven agents that don't need an internet connection.
How it squeezes memory down
Running a 5B or 8B model locally used to trigger aggressive app-killing on mobile OSes. Google DeepMind tackled this with a technique called Per-Layer Embeddings (PLE).
PLE significantly cuts RAM overhead during execution. As a result, the 5B model runs with a dynamic memory footprint of just 2GB, while the 8B version operates in around 3GB.
In practice, it delivers the capability of 5B and 8B architectures while consuming roughly the same memory budget normally required by 2B and 4B models. Google also co-engineered the runtime with Qualcomm, MediaTek, and Samsung System LSI to ensure smooth execution on modern mobile silicon.
Availability and rollout
Gemma 3n is available starting today in an early preview alongside technical documentation.
This same shared architecture serves as the foundation for the next-generation Gemini Nano, which Google plans to roll out across Android and Chrome features later this year.
Why it matters
For mobile developers and builders building local-first tools, Gemma 3n is a big deal. A 2GB dynamic memory footprint makes capable multimodal AI practical on mainstream smartphones, not just high-end flagship devices.
By releasing the same underlying architecture that powers Gemini Nano into the Gemma family, Google is effectively laying the plumbing for the next wave of on-device Android apps. It sets a new efficiency benchmark for edge AI and will push competitors to rethink their local inference architectures.
Sources (1)
- developers.googleblog.com(opens in a new tab)Originaldevelopers.googleblog.com