1. News
  2. Everyday tech
  3. Phones · Offline AI
Phones · Offline AI2 min read

A free Google app runs AI on a phone faster than one writer's PC

Google's free AI Edge Gallery app ran a model faster on a phone than Ollama did on one writer's older gaming PC, and it works offline.

By the Dashline deskSource: How-To Geek
A free Google app runs AI on a phone faster than one writer's PC
Photo: How-To Geek

A free Google app can run an AI chatbot on a phone faster than a fairly capable desktop PC. How-To Geek writer Arol Wright ran the same model, Gemma 4 E4B, on his phone and on his computer. The phone won by a wide margin. The app is called Google AI Edge Gallery, and it works without an internet connection.

What the app does

AI Edge Gallery is open source and available on the Play Store and the App Store. You pick an open model from a list, download it, and run it on the phone's own chip. Some models are hosted on Hugging Face and ask you to sign in and accept a license first. After that, no connection is needed, and prompts never leave the phone.

Beyond chat, the app can describe images and transcribe and translate audio recordings. With supported models, it can also show its step-by-step reasoning. Google still labels the app experimental.

Why the phone came out ahead

The test PC has a Ryzen 9 5900X processor, an Nvidia RTX 3070 graphics card and 32GB of RAM. Running the model through Ollama, a popular desktop tool, it took several seconds longer than the phone, and sometimes minutes longer. The article gives three reasons.

  • Memory speed: text generation depends on how fast a model's data moves through memory, and current phones have fast shared memory.
  • Graphics support: if Ollama cannot use the graphics card well, or the model does not fit in its memory, work falls back to the slower main processor.
  • Tuning: the phone models are packaged for Google's LiteRT runtime and shrunk to suit mobile chips.

What this does not mean

The writer is clear that phones have not overtaken PCs. The result says more about the age of his computer. A recent graphics card with 12GB or more of video memory would likely run the same model much faster than a phone. It could also handle far larger models that would never fit on a handset.

Limits to know

Each model download runs into the gigabytes, so storage matters. The app sometimes stalls while loading a model. Phones with less RAM are limited to the smaller models. The writer tested on a Google Pixel 9 Pro and an iPhone 18 Pro Max, which are recent flagship phones, so an older phone may not match his results.

What to do

Anyone curious about private, offline AI can try the app on the phone they already own. Start with a smaller model, check free storage first, and test it once in airplane mode to see how it handles everyday questions.

The takeaway

A recent phone can run private, offline AI surprisingly well, so try a small model in Google AI Edge Gallery before buying new hardware.

More articles

You should really read these.

Next step

Let’s talk