Private AI
for any
Android.

Zofi downloads quantized open-source models (GGUF format) optimized for mobile CPU inference. The model runs entirely inside the app using llama.cpp for Android — no cloud round-trip, no server. After a one-time model download over Wi-Fi, all inference is local and instant.

Any Android 9.0+ device with at least 4GB RAM. We use the Samsung Galaxy S10 (Exynos 9820) as the design baseline — if it runs smoothly there, it runs everywhere. Smaller models like Qwen3-0.6B work on 4GB devices. We explicitly test against mid-range hardware before each release.

Yes. Zofi is open-source (AGPL-3.0) and free at its core. There are no servers to fund during inference — your phone does 100% of the work. We may build optional premium services around storage or sync in the future, but the local inference app will always be free.

The app UI launches in Portuguese, English, French, Spanish, and Arabic. The underlying models (Qwen3, Gemma 4, Llama 3.2) natively support 30–50 languages including most African, European, and MENA languages. You can converse in any language those models understand.

The beta APK is available now — use the Download button at the top of this page. Google Play Store submission is our immediate next step. You can install the beta today — it is a complete, functional app, not a preview — and it will update automatically once the Play listing goes live.

AGPL-3.0 means anyone can inspect, modify, and build on Zofi — but derivative work must remain open. No silent corporate forks, no walled-garden takeovers. Technology that serves underserved communities should be accountable to those communities. That's the principle. That's why the license matters.