On-device AI
How We Run Google Gemma 4 Fully On-Device in Kukku
Today we're shipping Kukku 2.0, and with it a change we're genuinely proud of: your personal AI now runs on Google Gemma 4, executing entirely on your iPhone. No cloud round-trips, no account required to think — the model reasons on the device in your hand, and your day never leaves it.
Why on-device, and why Gemma
A personal AI sees the most personal data you have: your health and fitness, what you eat, your habits, and how you spend. The safest place for that data to be processed is the one place you already trust with it — your own phone. Gemma is what makes that practical. It's an open model family from Google DeepMind, small enough to run locally yet strong enough to power real agentic workflows, which is exactly the trade-off a private personal AI needs.
- Private by design. Inference happens locally; there is no cloud step to opt out of.
- Works offline. Put the phone in airplane mode and Kukku still answers.
- Low latency. No network hop means responses start right away.
- Open and inspectable. An open model is one you can reason about, not a black box behind an API.
What runs where
Kukku is local-first. Gemma 4 handles day-to-day reasoning directly on the device. For the occasional task that benefits from a larger model, you can connect your own remote or OpenAI-compatible endpoint with your own key — but that's your choice, off by default, and the private on-device path is always the default experience.
The default path never leaves the device. Remote escalation is opt-in and user-controlled.
What it powers
Gemma 4 is the engine behind Kukku's seeded agents — Health & Fitness, Food & Nutrition, Habits and Expenses — plus an Everyday agent and the ability to build your own. It reasons over your local context: snap a receipt and it logs the expense; photograph a meal and it estimates calories and macros; track habits and spending in per-agent journals. All of it on the phone.
Availability
Kukku 2.0 (Build 14) is Apple-approved and rolling out on the App Store on August 17, 2026. On-device inference with Gemma 4 is available on capable devices; feature availability can vary with device memory. Calorie, macro and health figures are estimates and not medical advice.
Attribution: Gemma is a family of open models built by Google DeepMind. Kukku uses Gemma under its model terms; "Google" and "Gemma" are trademarks of Google LLC, used here to describe the model that powers on-device inference.
Want to see it? Download Kukku Personal AI on iOS, turn on airplane mode, and ask it something. That's the whole pitch — private AI that runs where your data lives.