Introducing FunctionGemma 🤏270m model for function calling 📱can run in your phone, browser or other devices 🤖designed to be specialized for your own tasks

and the week is not over yet!

here's a TRL example for fine-tuning on mobile actions

To fine tune it for browser control using an RL environments directly check out this Colab:

Gemma 4 when

snooze fest, you really hyped us up for that? lmao

@mkurman88 look so smol model.

Gemma 4 model series release coming soon, right😿?

This is very cool. Somewhat like the “cognitive core”?

awesome! inference on mobile is still really limited based off model capability. interesting to see if better tool calling is enough to drive more mobile usage

this is a joke right?

fuking lovely man

Hi @osanseviero, that's great! Are you going to put it in the Google AI Edge Gallery? That way we can test it on our phones. By the way, congratulations to the Google AI team on their amazing week, month, and semester!

YESS 🚀🚀🚀🚀🚀🚀🚀

This is amazing Omar! I've been waiting for something like this for some projects of mine! Amazing job! Also on an unrelated topic, did you find out about the Angular issues with Gemini 3? (I was the guy asking about it on your latests Spanish AMA)

without finetuning this is junk, and only works on sample schemas. I tried with 10 tools, one of them is tap, and the model produces, this output, when it had to call tap I apologize, but I cannot assist with managing or logging in user accounts. My current tools are focused on managing app operations such as tracking tasks, launching applications, and managing user data. I cannot connect to or access user login credentials.<end_of_turn> <end_of_turn> <end_of_turn> <end_of_turn> <end_of_turn> and for Go back to the previous screen , it had to call back tool, it invented a new tool - mua_back_screen without finetuning not useful.

Just right when I needed it 💐 this is great

I wonder if I can use this with the home droid I'm working on 🤔

Love it

Let's goo , insane work 🔥🔥 Local LLMs are the future..

It's the last piece for to be truly private AI on edge

🔧

Honestly love this direction. Small, fast, does one thing well.

@osanseviero looks like example on ollama does not work: ResponseError: does not support tools (status code: 400)

Let's goo , insane work 🔥🔥 Local LLMs are the future..

Please make it apache

What's the lowest end iPhone that would be able to run a model like this?

insanely insanely cool

I've got a feeling this is going to be enormously useful for something. But I'm not sure what.

Gemma 4?

ok how to run it like the demo ?

Could we use this model to categorize incoming notifications into "important" and "not important" containers? And if not, please do it as Google yourself for your @GooglePixel_US lineup.

FunctionGemma's 270M parameter model enables function calling across various devices effectively.

Ok so this is the calm before storm