Tucano2-qwen-0.5B-Instruct: A Compact Portuguese LLM
Tucano2-qwen-0.5B-Instruct is an instruction-tuned Portuguese language model developed by Polygl0t, based on the Qwen3 Transformer architecture. Despite its compact size of 0.8 billion parameters and a 4,096 token context length, it demonstrates strong performance across various Portuguese benchmarks. The model was trained using a combination of Supervised Fine-Tuning (SFT) and Anchored Preference Optimization (APO), with all datasets, source code, and training recipes openly available for reproducibility.
Key Capabilities
- Retrieval-Augmented Generation: Effective for tasks requiring information retrieval from a given context.
- Function Calling and Tool Use: Capable of interacting with external tools and functions.
- Summarization: Generates concise summaries of text.
- Structured Output Generation: Produces responses in specified formats, such as JSON.
- Multilingual Focus: Primarily designed for high-quality interactions in Portuguese.
Good for
- Portuguese Language Research: Serves as a foundation for research and development in Portuguese NLP.
- Fine-tuning for Specific Applications: Adaptable for deployment in real-world applications under the Apache 2.0 license, with a recommendation for bias assessment.
- Resource-Constrained Environments: Its compact size makes it suitable for scenarios where larger models are impractical.
Performance Insights
While a smaller model, Tucano2-qwen-0.5B-Instruct shows competitive performance against other models in its size class on Portuguese benchmarks, particularly in Knowledge & Reasoning and Instruction Following, as detailed in the evaluation tables. For instance, it outperforms Qwen2.5-0.5B-Instruct across several metrics, including a notable lead in Knowledge & Reasoning (NPM) and Instruction Following.