The M1 Pro MacBook Pro can run Gemma 4 26B-A4B offline!

I bought a MacBook Pro M1 Pro with 16 GB of memory and a 512 GB SSD shortly after it was released. I think that was about four years ago? Not sure. It's August 2026 as of this writing, and it’s still my daily driver! I’ve used it to run the YouTube channel of a former client (Final Cut Pro, Apple Motion, Logic Pro). I’ve used it to edit photos in Lightroom, Photoshop, and Affinity. And I’ve used it to produce music using Logic Pro and other compute-expensive tools. I don’t remember when I last owned a Mac that kept running reasonably quickly for this long. Maybe never?

Anyways, I recently learned of a new project for Apple Silicon Macs that stunned me: Turbo Fieldfare. It’s an open source app to run LLMs larger than the physical memory of the device it’s running them on. It runs Gemma 4 26B-A4B, a ~15GB large model, offline, using about 2GB of memory! Its a very polished tech demo, clearly meant for people who know how Macs work and possess at least the basic knowledge of creating an app on Mac. To get it to run you’ll need to clone the GIT repository and build the app. There is no binary available you could just open and run. And once you’ve done that, you can download and run Gemma 4 26B-A4B locally, on your Mac from 2022.

On an aside: Since this app manages to run using about 2GB of memory, it could technically run on an iPhone 8 from 2017. Very slowly, but it should work, I think. I wonder how this tech evolves, because it would devalue the inference-data-centres large AI companies spent billions on massively, if running their models locally without paying for tokens became the new normal.