A field guide for Android engineers

Put the right model
in the right place.

Compare capable small models, understand the real device tradeoffs, and leave with a stack you can prototype—not just a leaderboard winner.

Open-source ≠ one license. This guide separates permissive models from open-weight models with custom terms.

local
private
fast
ON DEVICE
2B effective
networkoptional

01 · Stack builder

Make the constraints choose.

There is no universal “best” model. Choose the product constraints that matter, and the recommendation updates instantly.

Where should intelligence run?
Target phone memory 8 GB
Core capability
Optimize first for
Recommended build Strong fit

Private multimodal assistant

Keep prompts and media on the phone, with enough headroom for a mobile-first multimodal model.

APPKotlin + Compose
RUNTIMELiteRT-LM
MODELGemma 4 E2B
Why it fits

Native Android path, multimodal inputs, and strong quality for a phone-class model.

Watch for

“E2B” is effective compute—not download size. Test thermal behavior and peak memory on real target devices.

02 · The shortlist

Six models worth testing.

03 · Architecture patterns

Local when you can.
Hybrid when it counts.

01 / PRIVATE PATH

Everything sensitive stays in the APK boundary.

Best for personal notes, lightweight copilots, offline search, and features that must work on unreliable networks.

  • No per-token API bill
  • Predictable offline behavior
  • Device fragmentation becomes your problem
Android appCompose · Room
Local modelLiteRT-LM / JNI
private loop