Google's Gemma 4 Puts a Capable AI Model on Your Laptop

Google released the newest member of its open Gemma family on June 3, and the headline is not a benchmark score. It is the hardware requirement. Gemma 4 12B runs on a laptop with 16GB of memory, the kind of machine plenty of people already own, while performing close to models more than twice its size.

The model is open under the Apache 2.0 license, which lets developers download the weights, run them locally, fine-tune them and ship products without asking Google's permission. It is multimodal, taking text, images and audio as input. The audio part is the genuine first: Gemma 4 12B is the earliest mid-sized model to accept sound directly, rather than bolting a separate speech-to-text step onto the front.

What makes that possible is an unusual design. Most multimodal systems push images and audio through dedicated encoder networks before handing them to the language model. Gemma 4 drops those encoders. Images pass through a lightweight module that is little more than a single matrix multiplication, and raw 16 kHz audio is sliced into 40-millisecond frames and projected straight into the model. Fewer moving parts means a smaller memory footprint, which is how a 12-billion-parameter model squeezes onto consumer hardware.

Small model, large-model scores

On standard tests, Gemma 4 12B lands near Google's own 26B mixture-of-experts model while using less than half the memory. It scores about 77% on MMLU Pro, a broad knowledge and reasoning benchmark, and close to 79% on GPQA Diamond, a set of hard graduate-level science questions. Those are not frontier numbers, and they are not meant to be. They are striking because of where they run.

That is the real story, and it is one we have written about before. The idea that serious AI has to live in a distant data center is more habit than fact, and each capable small model chips away at it. A model on your own machine answers instantly, works offline, and never ships your data to anyone. For a doctor reviewing notes, a lawyer reading a contract, or anyone handling information they would rather not upload, that combination matters more than another point on a leaderboard.

There are limits worth keeping in view. A 12B model still trails the largest systems on the hardest reasoning, and running models locally puts the burden of updates and safety tuning on whoever downloads them. But the direction is hard to miss. Two years ago a model this capable needed a rack of servers. Now it fits in a backpack, and Google is giving it away.