Gemma 4 now works on-device using React Native! You can now run Gemma 4 fully offline in your cross-platform apps with local hardware acceleration: ⚡ Vulkan delegate on Android ⚡ MLX delegate on Apple Silicon See Gemma 4's vision and tool-use capabilities in action, instantly reading a flyer and... scheduling a calendar event, 100% on-device.show more

Shout out to @swmansion for support! Check out the opensource demo app: LLM docs:

Been shipping this exact pattern already - react-native-litert-lm, open source, Gemma running fully offline on RN/Expo with LiteRT-LM, and zero cloud calls.

Here is an example of high success using fine tune LoRA Gemma 4

Here is an example of Gemma 4 (LiteRT) deployed for local LLM generation for an agentic RAG use case.

And flutter ?

This makes adding private AI features much easier in cross-platform apps. The offline support will help with data security in many projects.

niiiice on-device gemma 4 :D which size?

Vision plus tool-use plus scheduling locally means the flyer-to-calendar demo works anywhere. That's the use case that finally makes on-device worth the model size tradeoff.

Are there research credits to help with fine-tuning? I only have a 3060

Pois....

@googlegemma can we use Skills with this integration? cc @swmansion

This is excellent. For devs, you'll likely need to request the "Increased Memory Limit" entitlement for iOS. Budget for slower reviews and approvals.

That is amazing!!!! Now I think it is also the time to look at the security side as well.

How much resources would it need

Haha what but don’t browser tabs cap ram usage??

This reminds me a funny correlation: - Gemma 4 works on React Native - LLMs don't run javascript while reading your apps - You are invisible to LLMs Check this out: (@isreadyai) 😉💚👇🏻

On-device offline support is a real win for cross platform apps.

Running Gemma 4 fully offline in cross-platform apps using React Native is a massive milestone for mobile dev. Vulkan on Android and MLX on Apple Silicon will ensure top-tier local hardware acceleration! 💻

100% on-device. 0% of your friends knowing you exist

Check this if you haven't:

Impressive update, @googlegemma! On-device AI with offline capabilities is a major step toward smarter, privacy-focused applications.

My friend, this is truly a monumental milestone in the evolution of accessible intelligence! 🚀 By bringing Gemma 4 to on-device capabilities through React Native, we are witnessing a beautiful shift toward the democratization of technology. The ability to run models fully offline, powered by local hardware acceleration like Vulkan on Android and MLX on Apple Silicon, is a testament to the incredible synergy between sophisticated engineering and practical, human-centric utility. Beyond the technical brilliance, there is a profound philosophical shift here: moving AI from the "distant cloud" to the "personal palm." It empowers developers to create more private, responsive, and seamless experiences, bridging the gap between high-level innovation and everyday human interaction. It is a significant step toward making powerful AI a portable companion that respects boundaries while pushing the limits of what we can achieve together. I am genuinely thrilled to see this level of accessibility unfold! ✨✨

I like that Gemma focus is on edge inference compared to other massive sized open models

I built a dllama fork for CPU tensor parallelism and am currently running Gemma 4 26b at 10.5tok/s on a raspberry pi cluster. This puts Gemma 4 models in edge AI device territory.

what's the cold start on mid range android, offline only helps after launch

Very impressive—bringing Gemma 4 fully on-device is a big step for mobile AI.

Does that include the app on ipad?

@swmansion When it comes to antigravity and replaces the gptoss😒

ew

Fully offline vision and tool use on device is a big deal, no more relying on a server round trip for stuff like reading a flyer.

the interesting part for react native teams is not just “ai on device”. it’s that more product behavior can live at the edge of the app instead of behind a round trip. that changes what feels instant, what works offline, and where you draw the native/server boundary.

Exciting!