Tucano2-qwen-3.7B-Instruct: A Specialized Portuguese LLM

Tucano2-qwen-3.7B-Instruct is a 3.76 billion parameter instruction-tuned model developed by Polygl0t, specifically designed for the Portuguese language. Built upon a Qwen3 Transformer architecture, it leverages a combination of supervised fine-tuning (SFT) and Anchored Preference Optimization (APO) using the Polygl0t/gigaverbo-v2-sft and Polygl0t/gigaverbo-v2-preferences datasets.

Key Capabilities

• Strong Portuguese Performance: Delivers robust results across various Portuguese benchmarks.
• Versatile Task Handling: Supports retrieval-augmented generation, function calling, tool use, summarization, and structured output generation.
• Open and Reproducible: All datasets, source code, and training recipes for the Tucano2 series are fully open and available.

Good for

• Portuguese Language Research: Serves as a foundational model for R&D in Portuguese NLP.
• Fine-tuning and Adaptation: Suitable for further fine-tuning for specific deployment scenarios under the Apache 2.0 license.
• Benchmarking: Offers competitive performance against other 3-4B parameter chat models in Portuguese, particularly in Knowledge & Reasoning tasks, where it achieves a Normalized Performance Metric (NPM) of 56.22, surpassing models like Jurema-7B and Qwen2.5-3B-Instruct in this category.