Google Just Made Their AI Free, Private, and Yours (Gemma 4)

Google has released Gemma 4, a family of open-weight models so lightweight that the smallest version can run on an $80 Raspberry Pi. The two-billion and four-billion parameter models fit on a phone or laptop and work entirely offline, with no subscription fees and no data leaving your device. The largest models in the family—26 billion and 31 billion parameters—are more powerful but still small enough to run on consumer hardware. And the performance is striking: these models would have been considered frontier intelligence just eight months ago.

What the Models Can Actually Do

The visual intelligence is the most immediately impressive capability. In one demo, a user holds up a banana in a cluttered room, and the model identifies not just the banana but the books on a shelf, the lamp, the curtains, and the room itself. This kind of multimodal understanding—text, images, and audio—is now available in a model small enough to run on a phone. One practical example: a Nest camera running Gemma 4 locally can identify a dog walking past the house or a delivery person carrying Amazon packages, and send a notification describing what it sees. That level of visual intelligence would normally cost significant money in API tokens from Claude or GPT, but here it runs for free on-device.

OpenClaw and the Cost Problem