📣 Call for contributions!

[sergiopaniego]

🌟 Open Call for Contributions: Gemma 3n Recipes 💎

Hello community! 👋

We got really excited last week with the full release of Gemma 3n 💎 by Google DeepMind. With that, we’d like to open this call for contributions to explore the amazing minimal recipes the community can build with it!

This project follows the same principles as our previous Hugging Face Llama Recipes, which you might also want to check out. You can refer to that repository to see the previous call for contributions and discover what the community created. You may also want to browse the Gemma Cookbook for further inspiration.

Star the project to boost it! ⭐

💡 What You Could Build

Gemma 3n is specifically designed for low-resource (on-device) environments and is a multimodal model that supports text, image, video, and audio inputs — so we’d love to see minimal recipes where those capabilities shine! ✨ We highly recommend checking out the blog post to better understand the possibilities and spark your creativity. Additionally, check out the already available recipes to understand what could be a good fit!

You could:

• Showcase agentic or interactive uses of the model
• Build recipes using transformers, llama.cpp, or transformers.js
• Create fine-tuning notebooks for domain-specific tasks using any of the supported modalities
• Or anything else that pushes the boundaries of on-device and multimodal AI! 🤖

Have your own idea?💡 Even better, we’d love to hear it!

📝 How to Contribute

1. Open a New Issue

   • If you’re interested in contributing, please open a new issue with a descriptive title.
   • Use the issue to outline your idea and gather early feedback.

2. Let Us Know

   • Comment on this post to let us know you’ve opened a new issue. We’ll update the main list with a link to your idea.

3. Start Building

   • Once you’ve shaped your plan, go ahead and implement your idea 👨‍💻.

4. Submit a Pull Request

   • When you’re ready, open a Pull Request (PR) that links back to your issue.
   • Mention the issue in your PR description so it's easy to track.

5. Update the README

   • Don’t forget to add a short description of your contribution to the README.md, so others can find and reuse it easily.

💬 For New Contributors

If you’re just starting out in open source, we recommend reading our Contribution Guide. It includes helpful tips to get you up and running.

We can’t wait to see what we’ll build together! 🤠

[AslauAlexandru]

@sergiopaniego I have some experience in fine-tuning LLMs such as BERT, GPT (free version) and so on. I use for fine-tuning LLMs unsloth ai and transformers (if I remember good). If helps for contributions. You can connect with me on linkedin link: https://www.linkedin.com/in/alexandru-aslău-328a3a301

[Sheetal-data]

Hi! I’ve opened an issue for a project called GemAgent- Gemma-Agent-a-smart-agent-built-on-Gemma-3n
It uses voice and text to set reminders, and guides users step-by-step through their tasks.

You can follow it here: #5
I’ve also started building it here: https://github.com/Sheetal-data/GemAgent-Gemma-Agent-a-smart-agent-built-on-Gemma-3n
Looking forward to contributing!

[sergiopaniego]

@sergiopaniego I have some experience in fine-tuning LLMs such as BERT, GPT (free version) and so so. I use for fine-tuning LLMs unsloth ai and transformers (if I remember good). If helps for contributions. You can connect with me on linkedin link: https://www.linkedin.com/in/alexandru-aslău-328a3a301

Thanks for the interest @AslauAlexandru!
Do you already have a specific idea or proposal in mind for a contribution?
Feel free to share it and we’d be happy to help you refine it! 🤗

[sergiopaniego]

Hi! I’ve opened an issue for a project called GemAgent- Gemma-Agent-a-smart-agent-built-on-Gemma-3n It uses voice and text to set reminders, and guides users step-by-step through their tasks.

You can follow it here: #5 I’ve also started building it here: https://github.com/Sheetal-data/GemAgent-Gemma-Agent-a-smart-agent-built-on-Gemma-3n Looking forward to contributing!

Super nice idea, @Sheetal-data and thanks for opening the corresponding issue!
This repo is particularly focused on minimal and self-contained recipes, so your project might be a bit broad in scope for this specific format.
That said, we’d love to help shape a smaller end-to-end version that could fit here!

We also have https://github.com/huggingface/cookbook, where a broader end-to-end use case could fit.

[Sheetal-data]

Thank you so much! @sergiopaniego I really appreciate the thoughtful feedback and the warm support.

You're absolutely right — the original idea is a bit broad for this repo. I’d love to shape a smaller, more focused version that aligns with the minimal, self-contained goals of gemma-3n-recipes.

I’d love any guidance on keeping it minimal but useful.

[yuvrajpant56]

Hey @sergiopaniego! 👋

I’ve just opened issue #6: Cookbook: Add “PDF Reading Assistant” RAG recipe. This notebook will walk through ingesting PDFs, embedding pages with SentenceTransformers, indexing in FAISS/Chroma, and serving a simple chat UI for QA.

Looking forward to any feedback or suggestions on how to shape it! 😊

[danielhanchen]

Hey guys! We made a Unsloth notebook for Gemma 3N 4B fine-tuning and multimodal inferencing in a free Colab Tesla t4! We also managed to fix multiple issues for it.

https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3N_(4B)-Conversational.ipynb

[yasir13001]

Hi @sergiopaniego 👋

I’ve just opened two Gemma 3n‑recipe proposals that show the model’s edge‑device super‑powers in very different domains:

---

🛒 Issue #7 – “ShopAI: Multimodal Order Extraction with Gemma 3n”

👉 #7 (comment)

• True multimodal demo — text 📝 + speech 🎙️ + vision 📷 in one recipe.
• Retail‑ready workflow — a casual request (“Add 2 kg sugar…”) or a product photo becomes structured order lines.
• Semantic inventory match — ChromaDB embeddings link items to the store catalog, showing live stock levels.
• Drop‑in backend — FastAPI endpoints for any voice assistant, chatbot, or web store.

---

♻️ Issue #9 – “SmartTrash: On‑Device Trash Classification & Sorting with Gemma 3n”

👉 #9 (comment)

• Edge‑AI sustainability — Gemma 3n vision model runs on a low‑power SBC to classify paper, metal, glass, plastic & other waste.
• Real‑time sorting — a camera‑equipped dust‑bin and simple servos route each item to the right compartment.
• Fully open source — code, embedded inference pipeline, CAD files, electronics schematics, and parts list will all be published.
• STEM‑friendly — perfect hands‑on project for schools and makers interested in AI + robotics + environment.

---

Would love any feedback on model fine‑tuning tricks, vision/audio pipeline optimizations, or ways to make these two recipes flagship showcases for on‑device Gemma 3n. Thanks for the inspiration, and looking forward to collaborating! 🚀

[koyeliaghosh]

Hi, I created a new issue -Real-Time Voice-Activated SOS System for Elderly Emergency Response Using Gemma 3n on Google AI Edge.

Enabling a real-time, voice-activated SOS system for elderly emergency response using Gemma 3n would have a significant impact, especially for vulnerable users in low-resource or connectivity-challenged environments.

Key strengths of this proposal:

Offline, on-device operation: Ensures reliability and privacy, even when internet access is unavailable.

Real-time voice recognition: Allows elderly users to call for help hands-free, which is critical in situations where they may not be able to reach a device or type.

Multilingual and accessible: Supports multiple languages and can be adapted for users with limited tech experience.

Immediate alerts and feedback: Notifies caregivers or emergency contacts and provides instant reassurance to the user.

This feature aligns perfectly with Gemma 3n’s mission to deliver accessible, life-saving AI assistance on affordable devices.
I strongly support this enhancement and look forward to seeing its development.
If there’s an opportunity to contribute or test, I’d be happy to help!

Offline, on-device operation: Ensures reliability and privacy, even when internet access is unavailable.

**Backend Architecture :**The backend to be built using Google AI Edge.
For real-time audio processing, use a lightweight server (e.g., FastAPI or Flask) to handle incoming audio streams, run speech-to-text inference, and trigger alert workflows.
Audio preprocessing (e.g., using ffmpeg or soundfile) to ensure audio is in the correct format for Gemma 3n.
Speech-to-text inference using Gemma 3n, running locally or on an edge device for privacy and low latency.
SOS keyword detection logic to identify emergency phrases.
Alerting system to send notifications (SMS, calls, app alerts) to caregivers or emergency services.
Optional text-to-speech for voice feedback to the user.
The backend can be deployed on local devices (phones, tablets, Raspberry Pi, or edge servers) to ensure full offline capability and data privacy.

[davidquicast]

Hi @sergiopaniego, I hope I’m still on time.
I’ve just published issue #16 . This example is based on the Gemma 3 Function Calling with HF document summarizer, which demonstrates function calling applied to text summarization.

My intention is to extend this approach to video summarization using function calling, leveraging Gemma 3n’s multimodal capabilities including audio, image, and text processing.

Let me know what you think.

[HyunZ118]

Hello @sergiopaniego,

First of all, I apologize for submitting a Pull Request before discussing it in this issue. I’m still new to contributing here and I missed the correct process outlined in this issue.

I’m a student focusing on medical AI with a deep interest in the Gemma model series. I prepared a notebook titled “Building MediAgent: A Medical Q&A Agent with Gemma-3n and LangChain.” Its goal is to demonstrate how to build an agent that helps patients better understand their medical discharge summaries.

Here is the PR link: #25

I would greatly appreciate any feedback on the implementation or structure, and I’m happy to revise and improve it based on your suggestions. If you’d prefer that I close the current PR and resubmit after further discussion here, I’m more than willing to follow that process as well.

Thank you for making this project open to contributors, and again, apologies for not following the process.