Google released Gemma 4 on April 2, 2026. It's their most capable open model yet — built on the same research as Gemini 3, released under Apache 2.0, and designed to run on everything from phones to data centers. The E2B variant is specifically built for mobile and edge devices.
Off Grid is a free, open-source app that runs Gemma 4 and other GGUF models entirely on your Android phone. No cloud. No account. No data leaving your device.
Which Gemma 4 Model Fits Your Phone
Text Generation
|
Attachments
|
Vision AI
|
Gemma 4 ships in four sizes. Two are practical for mobile:
Gemma 4 E2B (Q4_K_M) — ~1.3GB download. The edge variant. Designed specifically for phones and IoT devices. Around 2.3 billion effective parameters. 12 to 20 tokens per second on recent Snapdragon chips. Fits comfortably on 6GB RAM phones. This is Google's answer to "can a 2B model actually be useful?" and the answer is yes.
Gemma 4 E4B (Q4_K_M) — ~2.5GB download. Needs 8GB+ RAM. About 4 billion effective parameters. Noticeably better reasoning and output quality than E2B. 8 to 15 tokens per second on Snapdragon 8 Gen 3. The sweet spot for flagship phones.
The 26B MoE and 31B Dense variants are too large for phones but run well on desktop via Off Grid's macOS app.
Why Gemma 4
Built from Gemini 3 research. Gemma 4 inherits the architecture innovations from Google's flagship proprietary model. The intelligence-per-parameter ratio is the best of any open model.
Native multimodal. Gemma 4 processes text, images, and audio natively — not through bolted-on adapters. Vision and audio understanding are built into the model from training.
256K context window. Long conversations and full document analysis without losing context.
140+ languages. Broad multilingual support for global use.
Apache 2.0 license. Fully open. No usage restrictions. No license complications.
Benchmarks that matter. Gemma 4 E4B outperforms models with 2 to 3 times its parameter count on real-world coding and reasoning benchmarks. The 31B Dense model scores 84.3% on GPQA Diamond and 89.2% on AIME 2026 — elite company regardless of model size.
Setting Up Gemma 4 on Off Grid
- Install Off Grid from the Play Store
- Open the model browser — Gemma 4 variants appear in the recommended section
- Pick E2B for 6GB phones, E4B for 8GB+ phones
- Download over WiFi
- Start chatting
Off Grid automatically uses QNN NPU acceleration on Snapdragon 8 Gen 1+, Adreno GPU via OpenCL on older chips, or CPU on everything else. Switch KV cache to q4_0 in settings to roughly triple inference speed.
Privacy
Every Gemma 4 inference runs on your phone's processor. After the initial download, Off Grid makes zero network requests. Turn on airplane mode and verify. Everything works.
No analytics. No telemetry. No accounts. Open source and MIT licensed.
Off Grid also runs Qwen 3.5, Llama 3.2, Phi-4, image generation, vision AI, voice transcription, tool calling, and document analysis — all on device. Check the GitHub for the latest releases.
