During The Android Show today, Google shared more about Gemini Nano, including how it won’t be coming to the Pixel 8.
During the Q&A portion, a Google engineer on the Android generative AI team said that Gemini “Nano will not be coming to Pixel 8 because of some hardware limitations.”
Google is otherwise “working to bring Nano to more devices.” Specifically, more “high-end devices in the near future.”
This does not bode well for older/existing devices getting Gemini Nano. Specs-wise, the difference between the Pixel 8 and 8 Pro is RAM at 8 GB versus 12 GB. In terms of whether the Pixel 8 could run Gemini Nano at a technical level, it probably is capable as evidenced by how the entry-level Galaxy S24 also has 8 GB of RAM and uses it for on-device Magic Compose in Google Messages.
That said, running an intensive model — for a handful of features — might have other performance impacts on the rest of the user experience that Google does not think is worth making.
On the Pixel 8 Pro, Gemini Nano powers Recorder’s Summarize capability (more on that below) and on-device Gboard Smart Reply in WhatsApp, Line, and KakaoTalk.
Pixel 8 vs. 8 Pro vs. S24
Meanwhile, Google today detailed a bit more about how Gemini Nano powers the “Summarize” feature in the Pixel Recorder app. When Google started experimenting with such a capability in 2022, it was looking at cloud-based models:
User privacy is a top priority and we really wanted our most privacy-centric users, which for Recorder is roughly half of our users, to be able to use such a highly requested feature. This caused us to pause the project at the time.
Recorder then switched to Gemini Nano and a smaller team was able to implement it in 4 months:
We were pleasantly surprised with Gemini Nano’s reduced latency and with how capable the model was despite it being much smaller than what we used during prototyping.
If Gemini Nano is not coming to older Pixel devices, Google really should implement a cloud-based summarization for the other half of users that want this “highly-requested feature.” After all, Recorder already introduced a cloud-powered “Transcribe again” feature in December that makes possible speech-to-text in a lot more languages. Just make cloud summaries something users have to manually trigger for each recording and then delete that audio.
Update: Additional details and images added.
FTC: We use income earning auto affiliate links. More.
Comments