Swahili Gemma 1B: Specialized for English-to-Swahili Translation and Conversation

CraneAILabs/swahili-gemma-1b is a 1 billion parameter instruction-tuned model based on Gemma 3, developed by Crane AI Labs. Its primary focus is English-to-Swahili translation and Swahili conversational AI, accepting input in both languages but generating output exclusively in Swahili.

Key Capabilities & Performance

• Exceptional Efficiency: Achieves a BLEU score of 27.6 with only 1 billion parameters, demonstrating the highest BLEU-to-parameter ratio (27.6) among compared models.
• Outperforms Larger Models: Surpasses the Gemma 3 4B model (4x larger) by 153% in BLEU score on the FLORES-200 English→Swahili dataset.
• Competitive Quality: Delivers 94% of the performance of Gemma 3 27B with significantly fewer parameters, making it highly efficient.
• Core Functions: Beyond translation, it supports natural dialogue, summarization, creative writing, and question answering, all in Swahili.

Good for

• Language Learning: Ideal for practicing English-Swahili translation and improving Swahili dialogue skills.
• Content Localization: Efficiently translating materials into Swahili.
• Educational Tools: Developing Swahili learning assistants.
• Cultural Preservation: Creating and documenting Swahili content.
• Resource-Constrained Environments: Its small size allows efficient deployment on consumer hardware and mobile devices, with various GGUF quantizations available for optimized usage.