Breaking

Gemma 4 Gains Fully Offline On-Device Support in React Native

July 8, 2026 at 13:57 EDT

Foundation Models
Open Source
AI Agents

Google DeepMind's open-weight model "Gemma 4" can now run fully offline inside React Native apps. It supports local hardware acceleration via Vulkan on Android and MLX on Apple Silicon, allowing image recognition and tool use to run on-device without going through the cloud.

April 2, 2026 · Google DeepMind

Gemma 4 Runs Fully Offline Inside React Native Apps

The open-weight model now runs on-device with hardware acceleration — Vulkan on Android, MLX on Apple Silicon. A demo reads a flyer with the camera and schedules a calendar event, with no network connection at all.

network connection required — fully on-device

128K–256K

context length (small → larger models)

modalities on-device: text · image · audio

On-device memory footprint

Mobile-optimized variants vs. server-scale models — column height is proportional to GB required.

1.1GB

E2B
eff. 2B

2.5GB

E4B
eff. 4B

Server

12B / 26B MoE
active 4B

Server

31B
Dense

The offline demo flow

Camera reads flyer

→

On-device image recognition

→

Tool use schedules calendar event

Upside

Privacy — no data sent externally
Works fully offline
No API costs
Real-time responsiveness

Constraints

Files reach ~2.5GB — app bloat
Only a few tokens/sec, even with MLX
Heavy heating & battery drain
Memory leaks need optimization

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.