Google's Gemma 4 brings multimodal AI to 8GB phones

Google's Gemma 4 edge variants — E2B (2B params) and E4B (4B params) — are the first genuinely multimodal models designed to run locally on mainstream smartphones. At 4-bit quantization, E2B needs just 1–1.5 GB of memory; E4B tops out at 2–3 GB. Both fit comfortably inside the 8 GB RAM floor of most flagships since 2023, including iPhone 15 Pro and up.

The models handle vision, audio, text, and function calling — a capability combo that was rare even in models 10× larger just a year ago. E2B prioritizes speed (20–35 tokens/sec), E4B leans into reasoning and visual detail (12–20 tokens/sec). Pick your trade-off based on the task.

This is the threshold moment for on-device AI: not a stripped-down text stub, but a full multimodal stack that never leaves your pocket.

Why it matters: when capable AI runs locally on a sub-$400 phone, the cloud dependency — and the latency, privacy risk, and API cost that come with it — starts to look optional.