Gemini Nano, Google’s AI designed to run directly on Android smartphones, will finally be able to be present locally on the basic Pixel 8. Google no longer planned it due to hardware limitations, before backtracking.
Update from March 28 at 3:32 p.m.:
Google has changed its mind: finally, the base Pixel 8 will be able to run Gemini Nano, the less powerful version of Google’s LLM, Gemini, all locally. The manufacturer put forward a hardware limitation which prevented it from offering it, except that it also intended to bring it to the Pixel 8 Pro. This even though if the model “ Pro » is more powerful, it is equipped with the same chip as the Pixel 8.
Gemini Nano will therefore arrive on the Pixel 8 in the next update of the Pixel Experience, in a “ developer preview “, as indicated The Verge. Enough to provide summaries of audio recordings, or even intelligent responses on Gboard. If Google backtracks, we can think that it is under pressure from fans of the brand, disappointed by a promise that it would not have kept. Especially since the technical explanation did not seem to hold up: here again, we can predict that this is a desire by Google to segment its Pixel 8 range.
Original article from March 8, 2024:
Google has officially launched Gemini, its large language model (LLM) integrated directly into Android, but also into its entire software galaxy such as Google Docs and Drive. One of its three versions, Nano, is designed to run directly on smartphones, without using Google servers.
And it seems that the last flagship from Google, the Pixel 8, will not benefit from this, at least on the base model.
The non-pro Pixel 8 deprived of Gemini Nano
It was in Google’s official show for Android developers, The Android Show, that the news was confirmed by one of the manufacturer’s engineers. Gemini Nano « will not be available on the Pixel 8 due to some hardware limitations“.
We already knew that the Pixel 8 Pro could directly operate its AI directly on the device, reserving in particular advanced photo and video functions such as an improved Magic Eraser, but also the intelligent summary function of the Voice Recorder.
What hardware limitations?
But when it comes to LLM, we had every right to expect that a version would be compatible with the standard Pixel 8. The Tensor G3 chip being exactly the same in the two devices, they only differ through the quantity of RAM (8 versus 12 GB for the Pixel 8 Pro). If RAM is important for all inference tasks, one wonders what limits this model so much.
Google has in fact trained two models of Gemini Nano: Nano 1, for devices with low memory (1.8 billion parameters), and Nano 2, more RAM-intensive (3.25 billion parameters). We then wonder why the manufacturer would have thought of these two models if the less demanding of the two cannot run on the basic Pixel 8.
As the site reminds us9to5google, the first Samsung Galaxy S24 has a similar configuration and can run a local version of Magic Compose to help you write your SMS on Google Messages. On the Pixel 8, this feature will apparently work via the cloud.