A local Qwen AI app running offline on smartphones

A local AI app for smartphones has been released as an Android APK, with offline use as a main feature. It uses two Qwen 3 models: a 1.5B model and a 4B model, both heavily quantized to reduce memory needs.

The 4B model is aimed at phones with at least 12 GB of RAM. The smaller 1.5B model is aimed at mid-range phones and uses about 2.5 GB of RAM.

A bilingual LoRA is being built so the smaller model can use tools better, and the small model is also being improved through from a 32B model. The 4B model can work without LoRA, and a Windows version is planned soon.

Key points

  • The app runs a on a smartphone and can work offline.
  • It uses Qwen 3 models in 1.5B and 4B sizes.
  • The 4B version targets phones with at least 12 GB of RAM.
  • The 1.5B version uses about 2.5 GB of RAM for mid-range phones.
  • A bilingual LoRA and from a 32B model are being used to improve tool use.
Read original