**Every AI release claims to be “efficient now.”

Most of the time, that translates to:

still needs expensive hardware

still feels slow locally

still breaks on reasoning tasks

So when Google released Gemma 4 E2B, I honestly assumed it would be another lightweight model that looked good in benchmarks and failed in real usage.

I tested it anyway.

And after a week of running it locally, I think small models just crossed an important line.

My Setup

Nothing fancy.

ollama run gemma4:2b

Hardware:

MacBook Air M1

8GB RAM

Ollama

No external GPU

Performance I saw:

~40 tokens/sec average

First pull took around 3 minutes

RAM usage stayed around 5GB

Fan noise was surprisingly manageable

Most importantly:

it actually felt responsive enough to use continuously.

That’s rare for local models on weak hardware.

The Moment That Changed My Opinion

I tested a simple logic puzzle first.

The kind of question smaller models usually fail because they rush into an answer.

Without reasoning enabled:

wrong…