cYbEr tIMoR

There is no dedicated Ollama LLM trained specifically for Tetum (Tetun) yet. Because Tetum is a low-resource language, the best approach today is to start with a strong multilingual base model that can generalize across languages. Recommended Ollama models include Llama 3.1 (8B or 13B), Qwen (Qwen2.5/Qwen3), and Gemma 3, as they have good multilingual understanding and work well in cross-lingual scenarios.

For better Tetum performance, fine-tuning is strongly recommended. You can use LoRA or QLoRA to adapt a base model with Tetum text (news, government documents, education materials, or parallel Tetum–Portuguese/English data). This approach is cost-effective and practical, especially for local deployment in Timor-Leste, and aligns well with projects like chatbots, digital literacy tools, or government services.

If fine-tuning is not yet possible, you can still improve results using few-shot prompting and Retrieval-Augmented Generation (RAG). By providing Tetum examples in prompts or injecting relevant Tetum documents at query time, models like Llama 3.1 or Qwen can produce more accurate and context-aware Tetum responses. Overall, the best current strategy is multilingual base model + fine-tuning (or RAG) rather than training a Tetum LLM from scratch.