Este é uma versão quantizada em 4-bit no formato MLX do modelo Polygl0t/Tucano2-qwen-3.7B-Think, otimizada para inferência eficiente em dispositivos Apple Silicon.

Model Summary

Tucano2 is a family of open-source Portuguese language models developed by Polygl0t. This Think variant generates chain-of-thought reasoning traces enclosed in <think>/</think> tags before providing a final answer. It was fine-tuned using Supervised Fine-Tuning (SFT) and Anchored Preference Optimization (APO).

The 3.7B Think is the highest-scoring model in the Tucano2 family for knowledge and reasoning tasks, achieving a 54.07 NPM score — outperforming models like SmolLM3-3B (46.28) and Qwen3-4B (40.97) on Portuguese benchmarks.

Think vs. Instruct: The Think variant produces explicit step-by-step reasoning, making it suited for knowledge and reasoning tasks. For direct instruction-following and coding, use the Instruct variant instead.

Detail	Value
Original Model	Polygl0t/Tucano2-qwen-3.7B-Think
Base Architecture	Qwen3-4B
Parameters	3.76B
Context Length	4,096 tokens
Quantization	4-bit (MLX)
License	Apache 2.0

Resumo do Modelo

Tucano2 é uma família de modelos de linguagem de código aberto para o português, desenvolvida pelo Polygl0t. Esta variante Think gera raciocínio passo a passo (chain-of-thought) dentro de tags <think>/</think> antes de fornecer a resposta final. Foi ajustada usando SFT (Supervised Fine-Tuning) e APO (Anchored Preference Optimization).

O modelo 3.7B Think é o modelo com melhor desempenho da família Tucano2 para tarefas de conhecimento e raciocínio, alcançando 54.07 NPM — superando modelos como SmolLM3-3B (46.28) e Qwen3-4B (40.97) em benchmarks em português.

Think vs. Instruct: A variante Think produz raciocínio explícito, sendo mais indicada para tarefas de conhecimento e raciocínio. Para seguimento direto de instruções e programação, use a variante Instruct.

Benchmark Results

Benchmark	Score
ARC Challenge (5-shot)	55.38%
ENEM (3-shot)	71.52%
BLUEX (3-shot)	63.00%
OAB Exams (3-shot)	53.76%
BELEBELE (5-shot)	83.56%
MMLU (5-shot)	61.18%
GSM8K-PT (0-shot)	56.70%
IFEval-PT (0-shot)	31.67%
Total Avg.	51.27

Usage

Python (mlx-lm)

python

from mlx_lm import load, generate

model, tokenizer = load("pessini/Tucano2-qwen-3.7B-Think-MLX-4bit")

messages = [{"role": "user", "content": "Qual é a capital do Brasil?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

The model will output reasoning inside <think> tags followed by the final answer:

<think>
The user is asking about the capital of Brazil...
</think>

A capital do Brasil é Brasília.

Citation

bibtex

@misc{correa2026tucano2cool,
  title={{Tucano 2 Cool: Better Open Source LLMs for Portuguese}},
  author={Corrêa, Nicholas Kluge and Sen, Aniket and Fatimah, Shiza and Falk, Sophia and Landgraf, Lennard and Kastner, Julia and Flek, Lucie},
  year={2026},
  eprint={2603.03543},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2603.03543}
}

pessini/Tucano2-qwen-3.7B-Think-MLX-4bit

Tucano2-qwen-3.7B-Think (MLX 4-bit)