LiteRT (formerly TensorFlow Lite) optimized version of Swahili Gemma 1B - a fine-tuned Gemma 3 1B instruction model specialized for English-to-Swahili translation and Swahili conversational AI.

This repository contains MediaPipe task bundles optimized for mobile deployment on Android and iOS devices.

📊 Translation Performance

FLORES-200 Evaluation Results

Our Swahili Gemma 1B model demonstrates strong performance in English-to-Swahili translation:

Metric	Score	Ranking
BLEU	27.6	4th out of 5 models
chrF++	56.8	4th out of 5 models

Model Comparison

Model	Parameters	BLEU	chrF++	Efficiency*
Gemma 3 4B	4B	10.9	44.1	2.7
Swahili Gemma 1B	1B	27.6	56.8	27.6
Gemma 3 27B	27B	29.4	60.0	1.1
GPT-5 Mini	~8B	31.8	62.4	4.0
Gemini 2.0 Flash	Large	35.6	64.6	N/A

*Efficiency = BLEU Score / Parameters (in billions)

Key Performance Insights

🎯 Efficiency Leader: Achieves the highest BLEU-to-parameter ratio (27.6 BLEU per billion parameters)
🚀 Size Advantage: Outperforms Gemma 3 4B (4x larger) by 153% on BLEU score
💎 Competitive Quality: Achieves 94% of Gemma 3 27B performance with 27x fewer parameters
⚡ Practical Deployment: Runs efficiently on consumer hardware while maintaining quality

Evaluation Details

Dataset: FLORES-200 English→Swahili (1,012 translation pairs)
Metrics: BLEU (bilingual evaluation understudy) and chrF++ (character F-score)
Evaluation: Zero-shot translation performance
Model: swahili-gemma-1b-it-33000 checkpoint

📱 Available Models

File	Size	Quantization	Use Case
`swahili-gemma-1b-fp16-instruct.task`	~1.0GB	FP16	Recommended - Instruction following format with MediaPipe bundling
`swahili-gemma-1b-fp16-raw.tflite`	~1.0GB	FP16	Raw TFLite model - requires custom tokenizer integration

🚀 Quick Start

Android (MediaPipe)

kotlin

import com.google.mediapipe.tasks.genai.llminference.LlmInference

// Load the model
val options = LlmInference.LlmInferenceOptions.builder()
    .setModelPath("/path/to/swahili-gemma-1b-fp16-instruct.task")
    .build()

val llmInference = LlmInference.createFromOptions(context, options)

// Generate response
val response = llmInference.generateResponse("Translate to Swahili: Good morning")
println(response)

iOS (MediaPipe)

swift

import MediaPipeTasksGenAI

// Load the model
let options = LlmInference.Options()
options.modelPath = "/path/to/swahili-gemma-1b-fp16-instruct.task"

let llmInference = try LlmInference(options: options)

// Generate response
let response = try llmInference.generateResponse(inputText: "Translate to Swahili: Good morning")
print(response)

Web (MediaPipe)

javascript

import { LlmInference } from '@mediapipe/tasks-genai';

const llm = await LlmInference.createFromModelPath(
  '/path/to/swahili-gemma-1b-fp16-instruct.task'
);

const response = await llm.generateResponse('Translate to Swahili: Good morning');
console.log(response);

🌍 Language Capabilities

Input Languages: English + Swahili
Output Language: Swahili only
Primary Focus: English-to-Swahili translation and Swahili conversation

📊 Performance Metrics

Translation Quality (BLEU Scores)

Model	BLEU Score	chrF++
🥇 Swahili Gemma 1B	23.64	52.26
🥈 ChatGPT-4o-latest	[TBD]	[TBD]
🥉 Other Models	[TBD]	[TBD]

Evaluated on 1,012 English-to-Swahili translation samples.

📦 Model Variants Guide

1. `swahili-gemma-1b-fp16-instruct.task` (RECOMMENDED)

Best for: Most mobile applications - ready-to-use MediaPipe bundle

Input format: Natural instructions

Translate to Swahili: Hello, how are you?

MediaPipe formats as:

### Instruction:
Translate to Swahili: Hello, how are you?

### Response:

2. `swahili-gemma-1b-fp16-raw.tflite`

Best for: Custom integrations requiring direct TFLite model access

Requirements: You need to handle tokenization manually Use case: Advanced users who want to integrate with custom tokenizers or frameworks

🎯 Capabilities

Translation: English-to-Swahili translation
Conversational AI: Natural dialogue in Swahili
Summarization: Text summarization in Swahili
Writing: Creative and informational writing in Swahili
Question Answering: General knowledge responses in Swahili

💡 Generation Parameters

Optimal settings for mobile deployment:

javascript

// JavaScript/Web
const response = await llm.generateResponse(prompt, {
  temperature: 0.3,
  topK: 40,
  randomSeed: 42
});

kotlin

// Android
val response = llmInference.generateResponse(
    inputText = prompt,
    temperature = 0.3f,
    topK = 40,
    randomSeed = 42
)

swift

// iOS
let options = LlmInference.Options()
options.temperature = 0.3
options.topK = 40
options.randomSeed = 42

let response = try llmInference.generateResponse(
    inputText: prompt,
    options: options
)

📱 Mobile Integration

Memory Requirements

RAM: Minimum 3GB recommended for optimal performance
Storage: ~1.2GB per task bundle
CPU: ARMv8 or newer recommended

Performance Tips

Preload models during app initialization
Use appropriate quantization: FP16 provides good quality for mobile
Cache responses for repeated queries
Batch processing for multiple translations

🔗 Related Models

Original Model: CraneAILabs/swahili-gemma-1b - Full precision HuggingFace model
GGUF Quantizations: CraneAILabs/swahili-gemma-1b-GGUF - Optimized for llama.cpp/Ollama
Ollama: crane-ai-labs/swahili-gemma-1b - Ready-to-run with Ollama

🎨 Use Cases

Mobile Translation Apps: Offline English-Swahili translation
Language Learning: Practice Swahili with instant feedback
Cultural Apps: Create culturally aware Swahili content
Educational Tools: Swahili learning assistants for mobile
Offline AI: No internet required after model download
Edge Computing: Run AI locally on mobile devices

⚠️ Limitations

Language Output: Responds only in Swahili
Mobile Resources: Requires significant RAM and storage
Context Length: Optimized for shorter inputs on mobile
Quantization: FP16 requires more memory than INT4/INT8
Platform Support: Requires MediaPipe Tasks GenAI support

🛠️ Development Setup

Android

gradle

dependencies {
    implementation 'com.google.mediapipe:tasks-genai:latest.release'
}

iOS

swift

// Add to Package.swift
.package(url: "https://github.com/google/mediapipe", from: "0.10.0")

Web

npm install @mediapipe/tasks-genai

📄 License

This model is released under the Gemma Terms of Use. Please review the terms before use.

🙏 Acknowledgments

Google: For the Gemma 3 base model, support and guidance.
Community: For Swahili language resources and datasets
Gilbert Korir (Msingi AI, Nairobi, Kenya)
Alfred Malengo Kondoro (Hanyang University, Seoul, South Korea)

Citation

If you use these LiteRT models in your research or mobile applications, please cite:

bibtex

@misc{crane_ai_labs_2025,
    author    = {Bakunga Bronson and Kato Steven Mubiru and Lwanga Caleb and Gimei Alex and Kavuma Lameck and Roland Ganafa and Sibomana Glorry and Atuhaire Collins and JohnRoy Nangeso and Tukamushaba Catherine},
    title     = {Swahili Gemma: A Fine-tuned Gemma 3 1B Model for Swahili conversational AI},
    year      = {2025},
    url       = {https://huggingface.co/CraneAILabs/swahili-gemma-1b},
    organization = {Crane AI Labs}
}

Built with ❤️ by Crane AI Labs

Swahili Gemma - Your helpful Swahili AI companion, now on mobile!

CraneAILabs/swahili-gemma-1b-litert

Swahili Gemma 1B - LiteRT

📊 Translation Performance

FLORES-200 Evaluation Results

Model Comparison

Key Performance Insights

Evaluation Details

📱 Available Models

🚀 Quick Start

Android (MediaPipe)

iOS (MediaPipe)

Web (MediaPipe)

🌍 Language Capabilities

📊 Performance Metrics

Translation Quality (BLEU Scores)

📦 Model Variants Guide

1. swahili-gemma-1b-fp16-instruct.task (RECOMMENDED)

2. swahili-gemma-1b-fp16-raw.tflite

🎯 Capabilities

💡 Generation Parameters

📱 Mobile Integration

Memory Requirements

Performance Tips

🔗 Related Models

🎨 Use Cases

⚠️ Limitations

🛠️ Development Setup

Android

iOS

Web

📄 License

🙏 Acknowledgments

Citation

CraneAILabs/
swahili-gemma-1b-litert

1. `swahili-gemma-1b-fp16-instruct.task` (RECOMMENDED)

2. `swahili-gemma-1b-fp16-raw.tflite`