Yash-Booputh/gemma4-prep
Domain:
natural language processingRecord type:
projectsoftware
Creator:
Yas
Host:
Gemma 4 hackathon prep: 5 local AI capability projects (chatbot, vision, voice, RAG, agent) for Build with Gemma Mauritius 2026
# Gemma 4 Prep 🛠️
Prep projects for **Build with Gemma: Blue Mauritius & Beyond** (Middlesex University Mauritius, 29–30 July 2026).
Each folder is a **self-contained mini-project** demonstrating one capability of Google's
Gemma 4
open-weight model. Every project runs **locally** through Ollama — no API keys, no cost, works offline.
## Hardware target
Built and tested on an **HP Victus 16** — RTX 4050 (6 GB VRAM), Ryzen 7 7840HS, 16 GB RAM.
Model choice per project is tuned to fit this machine.
## Projects
| # | Project | Model | What it does | Status |
|---|---------|-------|--------------|--------|
| 01 | Chatbot | `gemma4:e4b` | Streaming text chat — the foundation UI others reuse | ✅ Working |
| 02 | Vision / Image Recognition | `gemma4:e4b`/`12b` | Describe, classify, OCR & answer questions about images | ⚠️ Blocked* |
| 03 | Voice Assistant | `gemma4:e4b` + Whisper | Full voice loop: speak → Gemma → spoken reply | ✅ Working |
| 04 | RAG — Chat with Documents | `gemma4:e4b` + nomic-embed | Upload PDFs and ask questions grounded in them | ✅ Working |
| 05 | Agent / Tool Use | `gemma4:e4b` | Gemma calls real tools (search, calculator, time) | ✅ Working |
*\*Vision code is complete, but Ollama 0.32.1 has a bug injecting images into `gemma4`
(the encoder runs, but the model doesn't "see" the image). See
02-vision-recognition/README.md for the diagnosis and fixes.
Everything else works today. See STATUS.md for the full setup report.*
## Gemma 4 model sizes (quick reference)
| Model | Runs on | Audio in? | Best for |
|-------|---------|-----------|----------|
| `gemma4:e2b` | phone / Pi | ✅ | ultra-light, offline |
| `gemma4:e4b` | laptop / 6 GB GPU | ✅ | **our default** — fast, fits, does audio |
| `gemma4:12b` | 8 GB+ | ✅ | best multimodal we can run locally |
| `gemma4:26b` | 16 GB+ GPU | ❌ | fast MoE, needs more RAM than we have |
| `gemma4:31b` | H100 / cloud | ❌ | max reasoning & fine-tuning (use Colab/Kaggle) |
> Gemma **outputs text only**. It reads …