google-gemma-3n-hackathon

[floral python]

Anyone having some confusion about the product to submit to this competition? We're asked to make a "private, offline-first" app, but it also requires a "Public Project Link (The Live Demo): A URL to your working product or interactive demo". Can this URL be the link to the app on the app store? If not, then how can an offline app be accessible thru an URL?

[queen aspen]

I am not able to run it locally on phone 🥹

It works on my pc with ollama though.

[celest herald]

My assumption is that the product should be capable of being a private, offline-first app, but we'll still need to show our work, and that is why it needs to be a link to a working product or interactive demo. As I develop a prototype, I plan on making sure something I do can run on my laptop but the code will be likely hosted on Github and include instructions for how to install and run locally.

[floral python]

Well so it is basically building 2 apps at the same time: a web app and a mobile app?
Can we get some authoritative clarification on this?

[burnt thunder]

The choice of the word “product”was to open it up for you to build anything, not only apps.

You can build robots, apps, websites, wearables, anything that uses Gemma 3n to make a positive impact on the world.

We will need to see the code to ensure your demo video is real.

[green wave]

Hi. Is there a deadline for joining the hackathon and team formation?

[fallow slate]

Has anyone been able to do audio on a pixel 9 phone

[foggy elbow]

Hi everyone

[mystic palm]

Hi folks!
I have a question.
Regarding selecting the track:
We have the overall and the special prizes.
Then in the special prizes section it says:
"Projects are eligible to win both a Grand Prize and a Special Technology Prize."
When i send the writeup, do I choose the special prizes and automatically will be in the running for the Overall?
Or... do I choose the Overall and automatically be in the running for a special prize? Which in my case will be the Ollama stuff?🤔
Thanks for any help.

[fast python]

Hey there where is Gemma- 3n linked

[glacial citrus]

Hi folks!
I have a question.
in 3rd problem statement. Its voice ton anlysis like by the ton of voice we have to detect the symptems or we just have to analyse the voice that what the particular user wants to say ? <@&1303433601177751593>

[fair estuary]

Wdym by linked?

[tranquil maple]

Did you ever get an answer on this?

[mystic palm]

At Kaggle discussion board, but I'm not sure I got it...

[hazy blade]

🚨 Collab Call: Gemma 3n + Education Hackers Wanted!

Hey folks — I’m Apreddy, solo builder of Lab2Life — a global hands-on science learning platform that just shook up Bolt Hackathon 2025. Now, I'm extending it to Gemma 3n (Kaggle Hackathon) to bring offline, privacy-first, multimodal science learning to underserved kids around the world.

Looking for 1 solid collaborator who’s got:
✅ Experience running Gemma 3n locally (preferably with Kaggle or on-device setup)
✅ Knowledge of multimodal AI (images, voice, text fusion)
✅ Comfort with Edge deployment, privacy-first AI, or education tech
✅ Bonus if you know lightweight backend/devops for packaging

💼 What You’ll Do:
Help build and demo a voice+vision AI Lab Assistant powered by Gemma 3n
Assist with integration (voice prompts, object recognition from webcam, etc.)

🎯 Goal: Build something powerful. Ethical. Fundable.
DM me or drop your GitHub/Kaggle/Portfolio if you're game.
Let’s bring science to every child — even offline.

[fair estuary]

Hey guys, didn't deepseek use the same thing basically?

[sullen shoal]

Hello, guys, is it possible to remove audio module to reduce memory consumption if my task doesn't require audio processing?

[rapid sandal]

robots would be crazy 🤣 but on a serious note, how would we go from writing code in the kaggle notebooks to using that for a full blown app? the two things just seem so separate—and maybe i didn't read the competition's instructions closely enough—but this really confuses me.

[burnt thunder]

You can use liteRT, Ollama, and other things to get your model running locally. I’d fine tune online, then deploy locally,

[burnt thunder]

Select Ollama.

[sweet swan]

Hello hello. I'm running Gemma 3n on my Android using the MediaPipe library. Sometimes my app will trigger an alert saying that the app stopped working, even though it's working fine. Has that happened to anybody here?

[restive frigate]

Can someone pls tell the model weight itself in 2-3gb then how the Gemma library app download size is just 110mb

[sweet swan]

Maybe the library comes without the models?

[sweet swan]

Another question I have is whether the PLE caching is enabled by default

[restive frigate]

Lmao yes found later

Thought would be insane genius of them to compress the models to that extent

[elder seal]

Due to latency issues that comes from running models on low end devices would it be termed cheating if the demos uses a rule based approach to ensure there is a smooth transition but the full version and all the integrations will be shown on GitHub. Note the models will be full optimized later to reduce latency, this cannot be done now due to time constraints and it will be clearly written in the Readme file

[restive frigate]

Is quantization allowed btw

[sweet swan]

What is a rule based approach?

[elder seal]

Using conditional statements like if apple == fruit return food. This approach is lightweight compared to the latency rate that will occured during demo presentation when using the actual models. Wanted to know if it's cheating to do that ?

[sweet swan]

Another important question could be if we need to show the actual latency in the demo or if the magic of edition can cover that

[elder seal]

It should cause it will take a miracle to run them on low end devices without quantization 😅

[elder seal]

That could also be another option but I think you have to clearly state it in the Readme file or in the demo cause you will still have to send a URL for them to test it themselves.

[late mango]

Has anyone tried build a webapp with Gemma-3n? I tried to build one and ran into memory issues although I'm running the app on RTX 3070. And then I found this issue.
https://github.com/google-ai-edge/mediapipe/issues/5976

[sweet swan]

Nope but I stumbled on that issue when I was looking into how to improve the memory consumption on my Android App using Gemma 3n

Are you using the Audio and Image capabilities?

[late mango]

I'm stuck at loading the model. The goal is to test with text generation and then explore audio capabilities.

[sweet swan]

Can we maybe build the product with a deployed LLM (Gemini) and use Gemma as a fallback? Of course that depends on the product, but is that against the rules?

[somber talon]

are all submissions automatically considered for the Overall prizes even if we choose special technology track?

On the competition page it is written " Projects are eligible to win both a Grand Prize and a Special Technology Prize." but on the writeup page we have the option to choose only one track. what does it mean?

[bitter wraith]

You can do multiple writeups or specify tracks in the writeup

[somber talon]

But we can submit only one writeups. I am asking about that single writeup. How are they considered?

[bitter wraith]

In this case, I think it’s best to submit this to the kaggle discussion forum of the challenge

[restive frigate]

is it true minicpm4 better than gemma3n in edge ai

[dusky star]

hey guys what options are using for GPU? my local model on mac is so slow to interpret images on time.

[restive frigate]

Kaggle notebooks online

[heavy ferry]

Hey all , can we use gemma 3n gguf quantized for this competition ?

[rapid umbra]

hey everyone, can anyone tell me if gemma3n on ollama support image and audio inputs? there is no clear communication if it does or not.
did anyone try it or not?

[restive frigate]

Yes

[lofty finch]

I've been pulling my hair out at this for a while lol, but has anyone successfully ran 2 different instances of gemma 3n at once on the dual kaggle gpus? Trying to increase inference throughput for testing purposes

[queen aspen]

What do you mean two different instances? With different weights? Why don't you just use batching?

[lofty finch]

Like load the model onto 2 different gpus at the same time

So I can use both for inference and double inference throughout

And I don’t mean split the model across the gpus*

[astral grail]

Hi, I wanted to ask if the registration for this competition is already closed? One of our teammates was hoping to join today, but she couldn't find the registration option on the competition page. Thank you!

[unkempt haven]

Use this page and start a write up and use the join hackathon off this page...

https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups

[vocal field]

You reckon an absolute coding beginner to attempt this challenge

I came up with an idea but I don't think it's very feasible in reality

Making ideas is easy, following through is not

I guess it's all about the learning anyway

I also have no team 😦

[vocal field]

I'll still try. Ignore my yap

Let me do my besttt

[surreal bramble]

No, the current version does not support it. Though it sounds like it's in progress.

[dusky star]

That's what I am worried about...even though it's stated that geamm3n supports multimodal functionalities i.e. text/images/audio, doesn't the current version only support text though? So how are people coming up with images/audio functionalities? Cuz technically they wouldn't really be using gemma?

[surreal bramble]

The model supports images, but ollama doesn't support it yet

[dusky star]

Oh...no wonder! So are the people using ollama at a disadvantage? if they can't really showcase the images/audio functionaities in their respective apps?

[restive frigate]

Have u tried using edge inference api itd the functionality in docs

[dusky star]

No, I haven't. I want to keep the install simple and local for now.

[restive frigate]

Hmm js make it work out when u shift to using inference api it must work so dw abt it now

[surreal bramble]

At the moment I'm using ollama and a regular gemma3 model with plans to switch when vision is supported

[rapid umbra]

atp, we can only expect that ollama will support audio and vision for gemma3n, or else we have to find workarounds for that.

hey all, if any info comes about ollama supporting audio and vision, pls let us know. it would be very helpful.

[lofty finch]

That moment when ai studio doesn’t even support non-text 💀

Kinda scuffed ngl

[dusky star]

What workarounds options would you guys be using?

[iron iris]

So we can develop also a web app, or we're just limited to mobile ?

[tall temple]

one of the evaluation metric Public Project Link (The Live Demo)

how to To publish app on App Store its 100$ and it should comply with all Apple compliance rules, the project being a prototype its highly unlikely. Is there any other way around.

[orchid tide]

Hey, all! Has anybody succesfully done fine-tuning in Unsloth and exported the model into a mediapipe task? I think i'm missing a step somewhere since the documentation in Unsloth only showed how to import as LoRA adapter, 16-bit, or GGUF. Sorry if it's a stupid question, I'm a beginner

[restive frigate]

Yess we can

[lofty finch]

Wait I’m actually so curious why Gemma 3n doesn’t have image support on the API

You would think that’s the first place that gets access…

[iron iris]

gemma3n doesnt have system prompr , tool calling , structured output, no multimodality in AI studio

[lofty finch]

😭

[lunar nacelle]

hello! does anyone know the extent of the code in which we need to open source? I know the model itself, the model training code, and a way to run it. do we need to open source the whole mobile app that we use in the video demo as well?

[dusky star]

Yes. Because that's how the judges will know if what you have coded accurately aligns with the video demo.

[fast python]

Hey everyone! I’m building a mobile app for the Gemma 3n Hackathon and had a few quick questions:
1. Is it okay to submit a mobile app as the main project?
2. Can it be iOS-only (built with Flutter)?
3. Are there any iOS version or device support requirements we should follow?
4. And is there a way to share the app with judges without having a $99 Apple Developer account?

Would really appreciate any help or official guidance — thank you!

[restive frigate]

Yes — there are workarounds:

Use TestFlight with a free Apple account (limited to 100 testers). You’d need to submit an IPA to Apple for review, which can take 24–48 hours.

Alternatively, provide an Xcode project or build scripts, and instructions for evaluators to run the app in Simulator or on-device via Xcode. Judges can then view functionality without needing to install an App Store or use a paid account.

[queen aspen]

Anyone interested in using capacitor? I wrote a plugin the lets you do inference through mediapipe on android and you can call it simply from javascript.

And it works both on mobile and desktop (through ollama on desktop)

[restive frigate]

And it supports audio?

[queen aspen]

I haven't added that yet but I think it won't be hard to add

I was thinking using gemma3n as an adhock speech transcription engine

There are some limitations in the mediapipe api. For example, you can only insert one image.

the docs for mediapipe llm inference are sparse, but their code is relatively readable

[fast python]

Thank you so so much!

[restive frigate]

Kindly add it then make it complete multimodally available

[tall temple]

Dont we need a developer account which os 100 USD per year?

[lost basin]

Hey everyone, I know the instructions said we needed a working prototype not just proof-of-concept in Kaggle notebooks. But is using Kaggle notebooks fine?

Because if I create a non-web (local) app, wouldn't the judges need to download and all. meaning I would need to add the model, or what?

What's fine for the live demo?

I would appreciate response and feedback, please. Thankyou

[lunar nacelle]

also I've noticed that for those of us using the google ai edge repo, is there an actual implementation for gemma 3n? I know the .task bundled model for the preview already exists, but for those of us wanting to fine tune it, there's no implementation for actually converting it to .task or litert format, be able to actually then run it locally right? or am I missing something

I see the gemma3 one, so one could use the text portion of the model since I believe they're similar enough, but then gemma3n also has other parts for audio and visual encoding which I do not see implemented.. anyone have any info on this?

[proven tusk]

Is it possible to enter multiple track with just one writeup?
How many writeups can we submit ??

[supple nest]

@mortal crow bot^^

[noble steeple]

I have the same question regarding write ups. My intention was to submit two projects - but it doesn’t seem clear to me at least.

[noble steeple]

www.kaggle.com/competitions/google-gemma-3n-hackathon/overview/submission-instructions

Pretty clear that this is a single-team, single-entry competition. Which is a bummer. Now I have to decide to dump one project 😢

[proven tusk]

But does one submission mean we can submit only to one track? Maybe we can submit the same project to two different tracks but it's one submission still?

[noble steeple]

If you go for the special tech tracks you’re still considered for the overall prize. I think if you state for overall, may not be considered for special tech tracks.

[proven tusk]

Is this mentioned somewhere ??

[noble steeple]

Yes it’s in the rules section while I was looking for the number of submissions.

[supple nest]

try mailing support or pinging a mod

[dusky star]

What are you guys submitting for the Live demo link if your project runs completely offline?

[noble steeple]

Simulate it in the browser. Like with this farmhandai.com live demo

[wanton finch]

does the autonomous robotics needs to be demo? that project promises alot of thing but the interactive demo really doesnt show much imo

[noble steeple]

This was a second project that I was intending to polish and submit but abandoned it.

According to the rules, a demo is nice to have. The way I had intended this farmhand ai was to showcase the open source shepherd rover - and how each of those special tracks could contribute to it.

I’d really focus on a track and showcase how Gemma 3n works- if you can get it working- great- if not, submit anyways.

[wanton finch]

oh ok! your project video was nice btw 🔥 yea im not too worried, have the pwa app ready and working, just gotta level up my story telling film now lol 😆

[wanton finch]

i think if you framed it as an economic equity and access to resources opportunities for small local farmers to compete with larger corporate farms. this would lead to a more resilience, and sustainable agricultural practices imo.

[noble steeple]

Thanks - yeah. I’m still going to open source the rover repo and dedicate some time to building out the docs and BoM. I think it could be worthwhile for those farmers

And I know there’s farmers who hack away at their tractors so this isn’t a really a skill problem- more of an organizing problem (open source hacking is amazing) and realizing that these small models can do a LOT.

If anybody is looking to takeover the farmhandai.com project which includes the shepherd rover - let me know. I can’t submit it as I have another project that I’m submitting.

[noble steeple]

Project was influenced by the problem that farmers have with determining when to best harvest their fields. Saw this on the show Clarksons Farm and thought about how local inference could help. Then thought about my experiences with the donkey car project.

[safe viper]

Any team that fully integrated gemma 3n to an android project (using kotlin)??

[urban merlin]

i have did it using dart flutter, its easy, but yeah its pain in ass to finetune and convert to .bin files , gave up finally

[dusky star]

114 submissions already 👀

[noble steeple]

Wow. Quite a bump over the weekend.

[toxic spade]

Awesome demo! I made and Android offline app. Wondering if it would be worth it to make a mock page on replit..

[noble steeple]

Thanks! Yes, I think adding a demo just to get the vision out before installing an app for users can help tell that story.

[toxic spade]

Yeah...I agree.. Thank you for your quick reply

[toxic spade]

If it's not too much to ask..what tools did you use for video? I am so not good with videos 😦

[noble steeple]

Google flow with veo 3. Start with a small storyboard with Gemini and break into smaller scenes for a more thorough flow veo prompt system. Store prompts in Google drive (as backup or future revisions) then feed it to flow for videos. Save videos as exports to Google drive and compose a new “Vid” document.

Edit within vids document by importing those saved clips from flow.

This is pretty non-linear but a system I felt comfortable with. Flow was ok with multi scene editing but not enough controls.

Google Plus subscription plus $25 in api credits to make two videos. Worth it.

[toxic spade]

Thank you sooo much !!!

[noble steeple]

Tip: use the “frames” to video option and use a still frame from a previous scene to help build the visual from the initial start of the new prompt

[safe viper]

we did use kotlin and it' hard af to integrate it

[toxic spade]

Great tip! I’ll try today this. I didn’t get great results last night.

[noble steeple]

if you're looking to run gemma 3n on edge devices: https://github.com/google-ai-edge/gallery/releases

[noble steeple]

yeah there's some odd results sometimes. I've done one particular scene about 10-15 times with different prompts and still not getting what I want accomplished.

[safe viper]

@drifting goblet @dark vessel @gray rapids @steel bay ??

[steel bay]

Thanks for the tag, public discords have been swamped with these bots lately and our automated detection doesn't catch them. New accounts seem to get hacked and spam us every few hours 😢

[safe viper]

that's sad to hear, i do hate when these bots does post these messages every where

did u ever tried developing a bot that maybe check how many messages were sent in everychannel ( calculate the frequency) then decide if the person is a bot or not??

[steel bay]

We might need to do something like that in the future if Discord doesn't solve the issue themselves

Generally we are protected a lot by kaggle account linking, but hacked accounts get through that

[safe viper]

from what i know, discord isn't gonna solve a thing😂

so i guess developping a solution might be the best thing

i do see that even with kaggle verification these ppl still spam, so maybe he isn't a bot but just a scammer

[steel bay]

It's always the same attachment based spam, so it's defintely hacked account driven by bots sadly

[safe viper]

i hope that this will be fixed asap

[dusky star]

Hey guys, for the writeup, did you write everything in the project description or upload a pdf file?

[noble steeple]

Project description and GitHub repository docs folder

[dusky star]

Just submitted my project 😭 . Can't wait to see what people have built. Good luck everyone. 🙌

[wanton finch]

ahh thank you! trying it now, the ingredient to video mode isnt available for me 🥲 was very cool on your video

[sullen shoal]

Hi! Did anyone manage to convert the model to .tflight for local running?? 😩 so little time left (( developed Android app, fine tuned the model, but no way to integrate them, please help!..

[fallen pollen]

Hi,

I’m stuck at 6 out of 7 tasks completed and need help with uploading an image in the Writeups section.

I’ve uploaded the image, but the progress indicator still shows 6/7 complete, and I’m unable to submit my project.

I’ve reviewed everything, but I’m not sure what might be missing. I’m feeling frustrated and would appreciate any tips on how to properly upload and embed the image so that it is recognized.

Thank you.

[queen gyro]

The Kaggle write up thing keeps telling me that the link to my youtube video is invalid. I kinda need to get past this ASAP since deadline is approaching. The link works fine b.t.w.

[queen gyro]

I am stuck also. Kaggle writeup keeps telling me the link to my youtube video is invalid. The link works fine everywhere else.

[fallen pollen]

It displayed a "Vertex AI resource exhausted" error and other unknown issues. I kept trying repeatedly and finally accepted the submission! 🌞

[shell aspen]

it working for me now

[queen aspen]

Is anyone else having trouble converting unsloth finetuned model to gguf?

It works in pytorch, but the moment I convert it to gguf, finetune goes away.

[lofty finch]

Unfortunately gguf and tflite conversion appears fucked

[queen aspen]

Well glad to hear I am not the only one facing this problem!

It's frustrating. They made base model gguf for ollama somehow, right?

So why can't I?

I even have baked the lora into the model correctly.

[hazy dune]

@everyone I made my app with Google AI studio it runs locally while submitting the code to the competition should we keep the api key or should we redact it and the competition judges will use their api keys

@drifting goblet @unique basalt @dark vessel @steel bay @marsh cairn

[noble steeple]

can anyone confirm - we can only submit one writeup per team? really like the secondary project I had lined up: farmhandai.com

[dusky star]

As per rule no 2: "SUBMISSION LIMITS
a. Each Team may submit one (1) Submission only."

[naive canopy]

goodmoring everyone 🙂 today is the last day, this is my first time, im so excited. goodluck to us 🥰

[lunar nacelle]

for the public link to demo, do you guys think a link to testflight to download the app will suffice or do I need to somehow host it on the web?

not sure how I would do the latter since its a mobile app

[toxic spade]

well, they say "if applicable" - This allows judges to experience your project firsthand, if applicable.
I wanted to do a presentation website, but the video is killing me..no more time left... you could replit your way into it if you have time
we also have a mobile app...

[timber compass]

did someone tried this notebook https://www.kaggle.com/code/valluruchetanreddy/gemma-3n-finetuning-vision-text-and-task ?
the "Convert Tensorflow to TFlite (liteRT)" part is not working for me

[noble steeple]

Thanks. I’m having a hard time letting go 😂

[runic falcon]

Hi Quick question - does flutter_gemma package qualify for the google ai edge track award ? since it wrapper model in a package also have mediapipe GenaAI. just wanted to confirm if it fits the criteria or if there's anything specific we should keep in mind. thanks in advance!

[sullen shoal]

Didn't work for me too. I fine tuned, pruned model and added final classification layer for more accurately fine tuning, completely removed audio module, so my model is pretty customized. I missed deadline, though was everything prepared, just left to integrate the model to android app that connects to desktop. I was so happy to get so many new things within short time: fine tuning/lora and pruning was my first experience. I'm still happy, it was a unforgettable adventure(got acc 0.92 with only 5 transformer layer/x6 shrinking) !

So, good luck everyone !

Let the

[rapid umbra]

Quick question everyone, I have built a desktop app that uses ollama. Now do I need a webapp also for the judges to try it out? Please guide as very little time is left.

Thanks

[pliant basin]

hey all, just wondering if we are allowed to share our gemma3n finetunes publicly (under gemma license) without authorisation or no? (wondering since it's gated on huggingface)

or do we need to provide some sort of gating mechanism so that only the judges can access it (or make people log in with HF)

[warm slate]

Same question. I also used ollama locally with streamlit, and now I might need to pay out of pocket to deploy it as a public link if I want the demo to be available to the judges to test out.

I was wondering if the github repository with instructions to setup locally, and a demo video would be enough, or do we need to host it publicly too ? Ollama makes things a bit complicated.

[pliant basin]

it does say "This allows judges to experience your project firsthand, if applicable", so maybe you can consider it not applicable (?). I'm in a similar boat - our team is working on an android app so obviously it can't be made to run on web

[dusky star]

This is what I did.

[noble steeple]

Since my flutter app uses a remote download for the default gemma3n model, I made a static website to document that (in addition to the main website) - so if you find it useful - great! https://livecaptionsxr-models.craigm26.workers.dev/

[noble steeple]

is it unnecessary? to a degree, yes, but it might be helpful to those who aren't as familiar with huggingface or kaggle and introducing models to a Flutter app.

[pliant basin]

out of interest, do you have any gating for it (i.e. user has to agree to gemma license) or no?

[noble steeple]

Darn great question. We do not have any gating. Probably should.

[pliant basin]

do you think it's required?

our group also is hosting a static site with a gemma model there etc

wondering if we should add a password - but this would make it a bit less accessible to people :(

[noble steeple]

Yeah. I think at minimum we need to cite the model source and other details so that it can be traced back accurately to the source.

I'm not sure if you need "gating" to that level. I'll dive into that area in a little bit. will report back what I find/do.

[pliant basin]

thanks

i'm just curious because the models are gated on huggingface

[noble steeple]

interesting. I pulled mine (version 1 LiteRT .task files) from kaggle: https://www.kaggle.com/models/google/gemma-3n/tfLite

[pliant basin]

huh

i cant attach photos in this channel but check

https://huggingface.co/google/gemma-3n-E4B-it-litert-preview

it's gated here

🤔

[pliant basin]

this also requires that you request access, though

[noble steeple]

that's right. I remember that process. pretty intense

[pliant basin]

if we do gate it, how do we ensure that only the judges have access 🤔

since the writeup is public, it's not like we can put a password in there

[noble steeple]

yeah. might have to lock it up after competition judging ends.

[pliant basin]

like, keep it un-gated for submission

[noble steeple]

that's probably the way to go at this point

[pliant basin]

just worried we would be violating some terms

https://ai.google.dev/gemma/terms#section-3: perhaps relevant

[noble steeple]

hah. same here. hopefully someone else can chime in. we're in the same boat. We want to make it easier for developers to re-submit to app stores and also download the base app without the huge model downloaded over and over. download model once and re-use it.

awesome find! - section 3.1

[pliant basin]

i mean, we can just add an 'agree to terms' button

[noble steeple]

yeah I added a bunch of text and some links

should be compliant.

[pliant basin]

let's hope so lol

theoretically someone can extract the URL and download it without reading the terms but eh

hopefully i'ts fine for a challenge

[dusky star]

375 submissions 👀

[toxic spade]

Recognition post for @noble steeple !
These past days I read syllable by syllable on how to create demo video with Google vids and Flow Veo3.
It’s not half as good as yours, but I made it before de deadline.

https://youtu.be/AC9Y_mH33aU?si=14fKixiOZIgLj1uO

Huge thanks! I couldn’t have done it without your help!!

[noble steeple]

I feel like there's going to some amazing submissions out of this. Multi-modality with offline is massive.

[noble steeple]

I already want to change up some of my previous projects that have computer vision and replace those with a customized per-project trained model. Problem with my submission with the flutter_gemma package (I worked with the package maintainer for this competition entry! Sasha Denisov) is that we ran into some on-device issues that’ll take more time to resolve than what we were able to submit today. Here’s our video: https://youtu.be/Oz8nzt2cc3Q?si=4uieI3oKNo44m-tU and website: https://livecaptionsxr.com

[toxic spade]

Cool idea.. I can see a use case for people that do not hear..or ADHD, elderly people

[noble steeple]

If we all sort of contribute to flutta_gemma, I think we can make our workflows a lot easier with multi-modal offline. https://pub.dev/packages/flutter_gemma

[toxic spade]

We tried.. but we got stuck at making the model tflite - https://github.com/nmo-genio/test_gemma3n_flutter

we evcen fine-tuned..but couldn't load the model correctly

I guess we can make server dedicated for this effort 😄

[noble steeple]

yeah common issue for multiple people here.

poll: if I create a discord server to help organize a community-effort develop the flutter - gemma Xn pipeline for ease of use - would you join? hit thumbs up

[stoic goblet]

craigm26 has been warned

Reason: Posted an invite

[noble steeple]

I can’t post the server link but I’ll post it in different places.

[toxic spade]

LinkedIn:)

[safe viper]

wish yall a good luck

[lofty finch]

I wonder how many more people could've submitted if we had day 0 support for multimodality in Ollama and other model serving tools, and conversions worked properly

Also, I'm kinda salty the latest transformers version broke generation with gemma 3n a day before the deadline

[safe viper]

i wonder how the comp would have gone if the audio modulatiy was available for mobile

[tranquil maple]

OMG that was tough but i made the deadline pivotted and started from zero 7 days ago

[safe viper]

good luck brother

[unborn nova]

Bro did anyone else have trouble saving it even though we pressed it in time before the deadline

It's probably just a lot of people are submitting but anyone

[safe viper]

it worked smoothly tbh

we submitted like 13 min before deadline

[unborn nova]

What should I do if I have screenshots lol that I submitted at around 4:57 PM pst so 3 minutes before

but nice @safe viper that's good to hear

[safe viper]

u didn't manage to submit at the end?

[unborn nova]

I pressed save but it wouldn't finish

[tranquil maple]

My only gripe is the lack of 4-bit quantization on the fine-tune. Did anyone figure that out?

[safe viper]

sad to hear, try to create a ticket and see if they could help

[lofty finch]

bro yeah, that was really annoying

[lunar nacelle]

wow so many good submissions. good luck everyone!

[unborn nova]

Hats off toe veryone the creativity was out of the box and everyone stepped it up with their edge computing

and creative hardware implemntations or usess

[naive canopy]

good morning everyone.. i see a lot of good stuff. im so happy 🥰

[untold cairn]

Hey!! I'm so sorry about this. I ran into some network issues right at the end and my submission for the hackathon was 4 minutes late. I worked really hard on it and would be incredibly grateful if it could still be considered. Thank you!

[noble steeple]

There’s a lot of great project submissions here

[tall temple]

Hey guys! Lemme know your views on my submission.
https://youtu.be/9jFIB0NQrnY?si=JheJqZj2my8teIwE

[toxic spade]

Biggest disappointment I had 😦
It was advertised mobile first … but…

[tall temple]

True

[toxic spade]

Awesome project! Congratulations!

[wanton finch]

yo thanks so much for this!
https://www.youtube.com/watch?v=tleK7c9nJCs&t
cooked this up just in time for the submission. im actually pretty surpised 😂

good luck everyone ✨

[toxic spade]

Yeah.. here https://www.kaggle.com/code/ionelmiroslav/gemma-3n-squad

[tall temple]

Thank you

[hollow crystal]

Hi, @burnt thunder Apparently My submission was only saved as a draft! All Items were submitted before the deadline, and my video was online on youtube as well. Can my writeup still make it in? I spent a lot of time making this system come to life and I thought the link kaggle made for me meant it had subitted:
https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/gemnet-emergency-mesh-network-with-gemma-ai-routin

here is the youtube video link that was posted an hour before the due time -
https://www.youtube.com/watch?v=qNSqGPjhF1Y

[sullen herald]

https://youtu.be/WdP8bOORbTs
let me know your thoughts on my submission

https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/bubble-stop-searching-start-doing

[restive vortex]

Good morning fam, I hope you are well and that everyone was able to make the submit, stay safe!

[pliant basin]

Our team had the exact same issue (couldn't convert unsloth to .task)

Got something that we hope is still good without fine-tuning though

[safe viper]

i mean we did found out the sad reality in a late stage

@steel bay sorry for the pin, but for competition like these, how much does it take to see the results

[toxic spade]

Yeah.. welcome to the power of marketing .. 😁

[safe viper]

dude we have a similar idea to this

[safe viper]

i mean next time i'll read be sure what the model can do before going deep into the idea 🥲

[safe viper]

btw what was your project's idea?

[toxic spade]

Originally? We had about 3 or 4 😂
First we wanted to buy drones to map a disaster field and proces video information with Gemma
Next we tried Flutter app
We ended up with android app for learning

There was Jetson idea at some point but the investment in hardware and learning curve was to steep

[tall temple]

We should have teamed up bro. Getting gemma 3n working with iOS was the hardest. It doesn’t support iOS gpu yet. “Mobile first AI”

[safe viper]

so we weren't the only one with crazy idea at first, i was trying to convince my team to buy raspberry pis, then we changed to a mobile app too

found out that our idea was impossible on IOS so we focused on android too lol

[safe viper]

even on Android it wasn't the easiest task i would say

Here is the video : https://youtu.be/f5UEdRCF6f4

[toxic spade]

you guys are crazy good at this video editing stuff!

[safe viper]

appreciate it, took me 2 days for scripting, editing and sfx

[safe viper]

was this your submission?

[noble steeple]

yes. had the farmhandai.com as a second project that I wanted to submit but re-reading the rules made it clear that we couldn't submit more than one project in a team.

I see tons of value out of all of these projects. created a new discord server to help bridge the gaps in our experiences and see if we condense all these down into a document.

local multimodal on common devices is very powerful

[safe viper]

i mean the one you submitet is very interesting ngl, well made man

btw i liked the server idea, a lot

[wanton finch]

if thats something you value, a pretty good idea would be an ai group discussion moderator so anyone can join and discuss topics in a space, you just need to prompt engineer the moderator as yourself to lead conversations. just a thought.

[wanton finch]

yea thats why i made an adaptive communication partner that can understand the full context of a human conversation—spoken words, text, and visual cues. this is the first time thats possible and can locally enforce data protection regulations like GDPR and HIPAA.

[toxic spade]

yup..same thing for me...2 days of video editing hell with zero experience

[safe viper]

would love to see the result tbh 😅

[toxic spade]

I previously shared 😅 https://youtu.be/AC9Y_mH33aU?si=vUa3WTf12ukQkTLM

[safe viper]

didn't see it before

but honestly u did a great job on the video, it looks great

and also the app idea is really good

[forest cliff]

Awesome submission guys!

Sharing mine as well, hope u guys like it 🙏

https://youtube.com/shorts/qO_A8_DId0g?si=GEZb4_WJw8hGYFSc

[restive vortex]

Good morning, very good project brother.. and excellent video! you got it right.

[noble steeple]

Summarized this challenge with our words in this discord above. Hopefully, it propels us towards working on easier deployments.

https://docs.google.com/document/d/1YhyzCXqx1AL7EQUEjgj4QT73hjl9Il_KKz9Muo7eCgQ/edit?usp=drivesdk

[forest cliff]

Thank you for the kind words brother 🙏

[forest cliff]

Require access permission to open sir

[noble steeple]

Updated.

If you want to help with moving multimodal offline inference- please DM me for discord server invite - it’ll be centered around gemma 3n and flutter, but id love conversations around the broader topic.

[restive vortex]

Awesome work, everyone! The submissions are super impressive. I've just shared my project too. Would love to hear your thoughts and see what you think of my findings!

https://www.youtube.com/watch?v=1Ofr-tD4PQw&t=43s

[pliant basin]

Does anyone know when the judging will be concluded?

[dusky star]

I’m guessing a week or two, maybe more, since there are about 600 submissions.

[noble steeple]

Weeks: https://www.kaggle.com/glennglen/discussion

[upbeat jacinth]

On Kaggle it said a few weeks

So many great project submissions, wow! So impressed and inspired that so many teams and individuals submitted projects for social good.

[safe viper]

a question for yall, we did see that many great project did come out in this short periode,
any team willing to continue developing their idea??

[unborn nova]

@safe viper yeah I'm probably gonna continue my idea, what about you?

[formal pelican]

I've been reading a few write-ups, your projects are amazing 👍 👍

[safe viper]

Depends on the team honestly. But the project is gonna continue for syre, if not now, then it's gonna after we got more experience for sure

[dusky star]

My project is a productivity and wellness app that I plan to use personally in my everyday life. As I use it, I will add more features based on its usability and accessibility. There is always room for further improvement.

[unborn nova]

Sounds great!

[safe viper]

sounds like a plan, i wish u a good luck

[formal pelican]

keren, mas 🇮🇩

Guys, I just wondering.
I put this hackathon project's code on Github. If we make some changes on our Github code while it's still in the judging process, is it gonna violate any rules?

[blissful mason]

Me.

Did any use Google AI edge specifically for llm inference? I tried using it but I am very new to Android development. Tbh it confused me a lot. There were lots of options.

Was anyone successful in using the npu on galaxy devices using the Qualcomm SDK ?

[forest cliff]

🇮🇩 thanks mas ilham

[pliant basin]

Yes

[blissful mason]

Do we have a spam bot or something. So we can block them ?

@pliant basin cool. If you don't mind what's your project ?

[pliant basin]

MAM-AI

https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/light-mam-ai

[noble steeple]

I’ve been taking a step back after this project ended and re-thinking about my desktop/workflow for work and side projects. Focused on local LLM/multimodal training and running inference - all locally.

Glad this project shifted my perspectives on cloud/local computes

[formal pelican]

Did anyone also use Gemma3n on Jetson Nano in this project? If so, how did you guys fine-tune it for specific case/dataset then run it on Jetson Nano? I know we can have Ollama to run it, but how about a fine-tuned Gemma3n? I have limited knowledge in this field.

[hushed creek]

Hi everyone 👋

Awesome submissions!

I’d like to share mine, SparkReader, an open-source ebook reader with a curated library of public domain books, which I made for the hackathon.

I’ve had the idea for a while, but the hackathon gave me the perfect incentive, deadline, and energy to finally create something meaningful. The vision is to connect us with the vast heritage of human thought (science, history, literature, etc), past and present, while embracing the potential of AI in an age when technology can sometimes distance us from it.

The app is fully open source, with the APK built automatically via GitHub Actions. It’s ready to go live on Google Play in a week or so, but adding more significant features will probably happen at my tortoise speed 🐢 (limited time).

If this project interests you, I’ve put together a CONTRIBUTING.md in case you’d like to jump in and help shape its future.

[wind gorge]

Hey everyone, thought this was pretty meta! Using my hackathon project to cram for my DAA test tomorrow.
https://media.discordapp.net/attachments/1406959867176161371/1406959982100086785/Screenshot_47.png?ex=68a45d17&is=68a30b97&hm=8ac00c1c6d39800c0755e1c5961db706a8a37f2c753787f0f3f18e38aa692631&=&format=webp&quality=lossless&width=966&height=543
https://media.discordapp.net/attachments/1406959867176161371/1406959982687158292/Screenshot_49.png?ex=68a45d17&is=68a30b97&hm=9120674acf837c402c60a4bd4c4febc95b94e52a60c77c0810cf110479dc9a46&=&format=webp&quality=lossless&width=966&height=543

Here is my writeup if you want to look into it: https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/aurenia

[kind saddle]

I see you are running the model on device like the Gemma 3n gallery app? I do the same.

[noble steeple]

Multimodal support with Gemma 3 nano now on the flutter_gemma package: https://pub.dev/packages/flutter_gemma

We’ll update LiveCaptionsXR after the judging is over.

[blissful mason]

Any idea when the judging would be done ?

[wind gorge]

My guess is around the start of september.

[noble steeple]

Gemma 3 270M fun: https://x.com/googleaidevs/status/1958242634108899622

[kind saddle]

I saw the announcement https://developers.googleblog.com/en/introducing-gemma-3-270m/ it could be a candidate for even a raspberry pi zero LLM on a stick. However I wonder how it would score on benchmarks the most against MS Mu 330M https://blogs.windows.com/windowsexperience/2025/06/23/introducing-mu-language-model-and-how-it-enabled-the-agent-in-windows-settings/

[noble steeple]

Yeah definitely. I think people are excited that the Gemma 3 model can be fine tuned and it would fit on those devices.

[wanton finch]

270m is wild. They need to train it with tool call ability. I’d make wonders out of that 😂 270M + mcp calls 🤔

[tall temple]

Does anyone know a good dataset to train model for tool call ability?

[wanton finch]

Training a model requires alot of resource. You might be thinking about fine-tuning a model instead. Here’s a few-shot example to get you started

Tool choices:

(a) ask for clarification
(b) web search
(c) bash command
(d) calculator
(e) respond directly

User: Hey.
Assistant: Hey! What's up?
User: What time is it
Tool choice: (c) bash command
bash command: date

User: Are you there?
Assistant: I'm here. What's up?
User: Can you order pizza
Tool choice: (a) ask for clarification
clarify: What kind of pizza?

User: Who's president?
Tool choice: (a) ask for clarification
clarify: You mean of the United States?

User: Who's president?
Assistant: You mean of the United States?
User: no, korea
Tool choice: (b) web search
web search: korea president

User: What's the meaning of life?
Tool choice: (a) ask for clarification
clarify: That's a philosophical question. Want me to see what the internet says?

Now you can run this on a top tier llm and pipe yourself a dataset. Just tell gpt5 or Gemini 2.5 pro to keep generating you as much of these as you need

[tall temple]

Thank you.

[pliant basin]

Fine-tuning is a type of training...

You may be thinking of pretraining when you are trying to draw that dichotomy

[wanton finch]

Great! Fun part about ai research is you can call it whatever you want. There’s no set rules! Perks of being first to a breakthrough

[wanton finch]

In this case your not teaching the agent what a bike is, but how to ride/use a bike. That’s how I think of it.

[pliant basin]

Hmm I disagree about that

There are definitely norms and conventions within the literature which usually are kept for the purpose of clarity

In this case fine tuning could be broadly considered as part of post-training (though sometimes reserved for the model developers themselves) which is still training because you're updating the model with data

[pliant basin]

Especially since training predates the current AI ecosystem of LLMs and its hype and you can also talk about training for example something seen more as a 'traditional statistical model' such as an HMM (I think that is where training as a term originated from - statistics and earlier machine learning). Broadly I feel like it's good to try keep the terminologies in line between the different fields 😄

But I guess that fine tuning is more precise when it comes to LLMs

[wanton finch]

Yup! however you understand it, that’s the way your gonna see it. You can be a follower or a visionary. Both is fine 😄

[zealous warren]

I’d hate to leave the conversation on an argument over semantics so here is my project 😄

🚀 Gemma Rover: A LeKiwi robot that uses Gemma 3n to make real-time decisions in a Mars-like environment
https://youtu.be/CN7uW8ERaWE

🔹 What it does

• Runs a continuous agentic control loop where Gemma 3n makes real-time decisions every few seconds.
• Integrates with LeRobot for navigation (QR-based) and manipulation (ACT policies).
• Demonstrates on-device autonomy for scenarios where cloud compute or human support isn’t available (e.g., a Mars mission).
• Scenario demo: rover collects soil → detects dust storm → seeks shelter → cleans solar panels → resumes mission.

🔹 How

• Gemma 3n runs locally on a Macbook M2
• 7 learned manipulation skills using LeRobot + ACT(also run locally on the Macbook)
• Open-sourced datasets + trained models on Hugging Face.

Writeup: https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/gemma-rover
Github: https://github.com/vladfatu/gemma-rover

[noble steeple]

I had a very similar secondary submission to this competition. It was a farm rover. Love the robotics approaches here.

I’d look at one of these https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/

[zealous warren]

with a price tag of $3499... I'll use the prize money from the hackathon if I win 😄

[noble steeple]

lol yeah. I'd probably get a NVIDIA DGX Spark and a MacBook Pro with some sort of Android XR setup. First submission project will be making updates after the competition.

[restive frigate]

nvidia dgx spark 4TB founders edition?

[noble steeple]

That’s the one

[restive frigate]

Have u already booked it?

[noble steeple]

just on the NVIDIA list and the ASUS version. Still debating if I actually need it for further model development or running local inferences on other projects. Or some sort of Enterprise deployment.

[restive frigate]

asus version noway the nvidia one should be better no 🤔 and can u do enterprise deployment on it 🤔

[main ingot]

hey everyone, quick question,when will the results be announced?

[noble steeple]

Probably soon

[noble steeple]

If you’re in SF: https://x.com/OfficialLoganK/status/1963756548552888601

[blissful mason]

Waiting eagerly.

[safe viper]

who's else check the results everyday like me 😅😅

[blissful mason]

Still no updates right ?

[safe viper]

to my knowledge, none

[bronze sapphire]

The fact that the judges lasted 3 months reviewing the submissions on the Gemma 2 competition (just notebooks) with about ~300 submission, talks about how much they're gonna last reviewing the 600 videos + technical writeup + demo/notebook

[safe viper]

so maybe results will be ready by 2026

[noble steeple]

30 plus hours of YouTube videos

[pliant basin]

Wow

Good point

[safe viper]

yeah also the fact that they'll review the code themselves

[manic crag]

Still I have an unpopular opinion .

Almost 2 months are there after competition end , Almost 60 days .

Don't you think like watching 60 hour of content in 60 days for 4 5 member team is too difficult? I don't think so . But it's like now I am losing my pascience .

I think they are going through the code and ig test and all take a lot of time . Or for winners might be they had some stages .

Anyway waiting for that one email or notification 🙂

I would like to appreciate everyone, you guys are so creative ✨️

[safe viper]

tbh, i hope that the results come soon, excited to see who won ngl

[formal pelican]

I'm excited as everyone here, but ..

https://www.kaggle.com/competitions/gemma-language-tuning/overview
The last Gemma 2 hackathon took about three months to announce the winners

Hopefully, this one won’t take longer

[manic crag]

That's too long time tho .

We need to wait too much 🙃

[restive vortex]

Hello fam, i hope all are good 💯

[safe viper]

yeah, doing alright

how about yourself dude?

[restive vortex]

I'm glad you're okay, everything's fine here, it's hot as hell. ☀️

[safe viper]

thanks dude

may i ask you where is that

because the weather hear is nice

or maybe i'm just the one who don't go outside a lot to know if it's hot or not lol

[restive vortex]

Venezuela

[safe viper]

nice to meet you dude