→ Live Demo on Hugging Face Spaces

Experience Gemma-3n-Swahili instantly in your browser - no downloads, no setup, just click and chat in Swahili!

🎥 Click here to watch the phone demo on YouTube

1. Download PocketPal AI app

   • 📱 iOS App Store
   • 🤖 Google Play Store

2. Search for the model in the app:

   Nadhari/gemma-3n-swahili-E2B-it-gguf
   

3. Download and start chatting - It's that simple!

ollama run hf.co/Nadhari/gemma-3n-swahili-E2B-it-gguf:Q8_0

For a better chat interface, use Ollamate - check their repository for detailed installation instructions.

Despite being spoken by over 200 million people across Sub-Saharan Africa, current AI models have limited Swahili comprehension and instruction following, leading to:

• Inability to maintain cultural understanding
• Limited support for code-switching common in modern Swahili usage
• Lack of specialized vocabulary for emerging technological concepts

Gemma-3n-Swahili represents a paradigm shift in how we approach low-resource language modeling. We've created models that excel specifically in Swahili while maintaining the ability to bridge to other languages when needed.

Explore our complete model collection on Hugging Face:

→ Gemma-3n-Swahili Collection

• gemma-3n-swahili-E2B-it: Balanced performance for resource-constrained environments
• gemma-3n-swahili-E4B-it: Optimal for general Swahili applications
• GGUF Versions: Quantized models for mobile and edge deployment

Our models achieve significant improvements over base Gemma 3n on Swahili benchmarks:

These results demonstrate transformative gains in cross-lingual understanding.

We fine-tuned Gemma-3n models using:

• Dataset: 10,000 high-quality Swahili instruction-response pairs from Bactrian-X
• Method: LoRA fine-tuning with Unsloth framework
• Focus: Language-specific improvements while preserving multimodal capabilities

• ✅ Excellent Swahili understanding and generation
• ✅ Technical and creative writing capabilities
• ✅ Cultural context awareness
• ✅ Code-switching support
• ✅ On-device deployment ready

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Nadhari/gemma-3n-swahili-E2B-it")
tokenizer = AutoTokenizer.from_pretrained("Nadhari/gemma-3n-swahili-E2B-it")

# Example usage
prompt = "Eleza kwa nini anga ni bluu?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

We welcome contributions!

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

• The Bactrian-X team for the Swahili dataset

For questions, suggestions, or collaborations:

• Open an issue on GitHub
• Reach out on Hugging Face