Nano A Nono: Pixel 8 Phones Too Dumb For Google's Smallest Gemini AI Model

Nano, Google's smallest AI model in its generative Gemini series, will not be available on Pixel 8 handsets due to "some hardware limitations."

Announced in December, Nano runs on high-end Android smartphones such as Google's latest Pixel 8 Pro and Samsung's Galaxy S24 line. The model provides various typical AI features of the day on those devices, such as summarizing audio recordings as text and Magic Compose, which can automatically generate quick replies to text messages for you.

These capabilities will not be coming to the Pixel 8, though, which like the Pro went on sale in October. In a panel discussion during the latest episode of The Android Show, Terence Zhang, an engineer working on generative AI on Google Android, said this was down to "some hardware limitations." (Skip to about 42:57 in the video below.)

Youtube Video

It's not clear what those limitations are exactly, considering both the Pixel 8 and Pixel 8 Pro devices are equipped with the same Google Tensor G3 processor and CPU and GPU cores. The Pixel 8 Pro, crucially, has more memory. The 8 has 8GB of RAM and the 8 Pro has 12GB, and by the sounds of it, that 8GB will cramp the performance of the model. So, no Nano for the Pixel 8.

Google declined to comment on the record.

• The perfect workspace collaborator
• Stack Overflow to charge LLM developers for access to its coding content
• Google to reboot Gemini image gen in a few weeks after that anti-White race row
• Google releases Gemma – LLMs small enough to run on your computer

Reasonably competent large language models are typically memory and compute intensive, making them tricky to run on smartphones, even ones with 8GB of DRAM.

Google developed AICore, an Android system service applications can use to make Google Nano perform ML inference tasks, such as text generation, on the device using whatever hardware acceleration is to hand.

Zhang said AICore is basically an API surface for Gemini Nano, and that this model is constrained by the computing power of the device. On phones where there is no Nano, such as the Pixel 8, there is no AICore to use.

AICore is useful in that it allows apps to tap into Gemini Nano on the local device without needing an internet connection to some remote cloud service. All the text generated by the model is private since the data doesn't leave the smartphone, which also reduces latency, making it faster to run. "Lastly, and most importantly, it's free," Zhang said. Since all the workloads are run using the hardware on devices, it doesn't cost a dime to interact with the model, unlike accessing an LLM from the cloud.

"AICore is currently only available on Google Pixel 8 Pro and Samsung S24 Series devices," Google makes clear in the developer documentation.

The web giant is said to be planning to roll out more AI features for its voice-controlled Assistant, powered by its larger Gemini Pro and Ultra models, next year on its Pixel devices. ®