Our Models

State-of-the-art language models fine-tuned for African languages, optimized for efficiency and cultural relevance.

Swahili Gemma 1B

Fine-tuned Gemma 3 1B Instruction Model

HuggingFace Ollama

Overview

A fine-tuned Gemma 3 1B instruction model specialized for English-to-Swahili translation and Swahili conversational AI. Accepts input in both English and Swahili, but outputs responses exclusively in Swahili.

Performance Metrics

BLEU Score 27.6
chrF++ Score 56.8
Efficiency Ratio 27.6 BLEU/B
Context Length 4,096 tokens

Capabilities

  • • English-to-Swahili translation
  • • Swahili conversational AI
  • • Text summarization in Swahili
  • • Question answering in Swahili
  • • Creative writing in Swahili

Key Achievements

  • ✓ Highest BLEU-to-parameter ratio among comparable models
  • ✓ Outperforms Gemma 3 4B (4x larger) by 153% on BLEU score
  • ✓ Achieves 94% of Gemma 3 27B performance with 27x fewer parameters

Usage Example

from transformers import pipeline

# Load the model
pipe = pipeline("text-generation",
                model="CraneAILabs/swahili-gemma-1b")

# Translate English to Swahili
result = pipe("Translate to Swahili: Hello, how are you?")
print(result[0]['generated_text'])

Ganda Gemma 1B

Fine-tuned Gemma 3 1B Instruction Model

HuggingFace Ollama

Overview

A fine-tuned Gemma 3 1B instruction model specialized for English-to-Luganda translation and Luganda conversational AI. Accepts input in both English and Luganda, but outputs responses exclusively in Luganda.

Performance Metrics

BLEU Score 6.99
chrF++ Score 40.32
Efficiency Ratio 6.99 BLEU/B
Context Length 2,048 tokens

Capabilities

  • • English-to-Luganda translation
  • • Luganda conversational AI
  • • Text summarization in Luganda
  • • Question answering in Luganda
  • • Luganda language research

Key Achievements

  • ✓ Highest efficiency ratio among compared models
  • ✓ Outperforms Gemma 3 4B (4x larger) by 535% on BLEU score
  • ✓ Outperforms GPT-4 Mini by 36% on BLEU score

Usage Example

from transformers import pipeline

# Load the model
pipe = pipeline("text-generation",
                model="CraneAILabs/ganda-gemma-1b")

# Translate English to Luganda
result = pipe("Translate to Luganda: Hello, how are you?")
print(result[0]['generated_text'])

Model Variants

GGUF Format

Optimized for CPU inference with llama.cpp

• Swahili Gemma 1B-GGUF
• Ganda Gemma 1B-GGUF

LiteRT Format

Optimized for on-device mobile inference

• Swahili Gemma 1B-litert
• Ganda Gemma 1B-litert

MLX Format

Optimized for Apple Silicon (M-series chips)

• Swahili Gemma 1B-mlx-fp16