Cactus Gemma

A multimodal mobile app powered by the Cactus inference engine, running Gemma 4 E2B - vision, speech, and text - on-device. Point your camera, ask out loud or type, get a streamed answer.

Features

• Snap & ask - take a photo, speak or type a question, watch a streamed markdown reply.
• Live mode - hold to ask; the model captures a frame and your voice prompt on each press.
• On-device first - runs locally with automatic cloud handoff when the local model's confidence falls below threshold.

Quickstart

Requirements: Node 20.19.4+, Yarn, and Xcode (iOS) or Android Studio (Android). A physical android/ios device.

git clone https://github.com/cactus-compute/cactus-gemma4.git
cd cactus-gemma4
yarn install

# iOS
yarn expo run:ios

# Android
yarn expo run:android

On first launch the app downloads 4-bit weights (~5 GB) from huggingface.co/Cactus-Compute/gemma-4-E2B-it. Apple devices pull the NPU-optimised build; Android pulls the CPU build.

iOS signing: app.json ships with a Cactus Compute appleTeamId. If you're building from outside that team, replace expo.ios.appleTeamId with your own Apple Developer Team ID.

References

• Cactus - the on-device inference engine that powers this app
• Gemma - Google's open model family
• Cactus-Compute/gemma-4-E2B-it - 4-bit weights used at runtime