Tucano2-qwen-0.5B-Base is a decoder-only transformer continually pretrained from Qwen3-0.6B-Base. Tucano2 is part of the Polygl0t initiative, which aims to advance language models for low-resource languages.

Tucano2-qwen-0.5B-Base shares the same tokenizer as Tucano2-0.6B-Base. Token embedding transplantation via Orthogonal Matching Pursuit was used to adapt Qwen3-0.6B-Base to be more sensitive to the lexical, morphological, and orthographic properties of Portuguese.

The model was continually pretrained on approximately 50 billion tokens and achieves state-of-the-art performance across several benchmarks designed to evaluate Portuguese language models. All data, source code, and recipes used to develop the Tucano2 series are open and fully reproducible.

Details

Architecture: a Transformer-based model (qwen3)
Size: 490,799,104 parameters
Context length: 4,096 tokens
Dataset(s):
- Polygl0t/gigaverbo-v2
- Polygl0t/gigaverbo-v2-synth
Language(s): Portuguese
Batch size: 1,048,576 tokens
Number of steps: 50,000
GPU: 8 NVIDIA A100-SXM4-80GB
Training time: ~ 59 hours
Emissions: 86 KgCO2 (Germany)
Total energy consumption: 225 kWh

This repository has the source code used to train this model. The full configuration used for training is available in the following config file:

Single stage (linear warmup with cosine decay): training_config.yaml

Checkpoints

Checkpoints were saved every 2,500 steps, which equates to approximately 2.5 billion tokens. The main branch of this repository contains the final checkpoint saved at step 50000. All other checkpoints are available as separate branches. To load a specific checkpoint, you can use the following code snippet:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Polygl0t/Tucano2-qwen-0.5B-Base"
revision = "step-2500"  # Change this to the desired checkpoint branch
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, revision=revision)

Or, you can access all the revisions for the models via the following code snippet:

python

from huggingface_hub import list_repo_refs
out = list_repo_refs("Polygl0t/Tucano2-qwen-0.5B-Base")
branches = [b.name for b in out.branches]
print(branches)

Learning Curves

This plot illustrates the evolution of model performance (measured by loss) as a function of training time, measured in tokens seen during training

Gradient Norms (L2)

This plot illustrates the evolution of gradient norms as a function of training time, measured in tokens seen during training.

Intended Uses

The primary intended use of Tucano2-qwen-0.5B-Base is to serve as a foundation for research and development involving Portuguese language modeling. Checkpoints saved during training are designed to provide a controlled setting for performing comparative experiments, specifically regarding the effects of continual pretraining on the performance of currently available benchmarks. You may also fine-tune and adapt Tucano2-qwen-0.5B-Base for deployment if your use follows the Apache 2.0 license. If you decide to use Tucano2-qwen-0.5B-Base as a basis for your fine-tuned model, please conduct your own risk and bias assessment.

Out-of-scope Use

Tucano2-qwen-0.5B-Base is not intended for deployment. It is not an out-of-the-box product and should not be used for human-facing interactions.
Tucano2-qwen-0.5B-Base is for the Portuguese language only and is unsuitable for text generation tasks in other languages.
Tucano2-qwen-0.5B-Base has not been fine-tuned for downstream tasks.

Basic usage

python

from transformers import GenerationConfig, TextGenerationPipeline, AutoTokenizer, AutoModelForCausalLM
import torch

# Specify the model and tokenizer
model_id = "Polygl0t/Tucano2-qwen-0.5B-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

# Specify the generation parameters as you like
generation_config = GenerationConfig(
    **{
    "do_sample": True,
    "max_new_tokens": 150,
    "renormalize_logits": True,
    "repetition_penalty": 1.2,
    "temperature": 0.1,
    "top_k": 50,
    "top_p": 1.0,
    "use_cache": True,
  }
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
generator = TextGenerationPipeline(model=model, task="text-generation", tokenizer=tokenizer, device=device)

# Generate text
prompt = "# A floresta da Amazônia: um lugar de Magia\n\n"
completion = generator(prompt, generation_config=generation_config)
print(completion[0]['generated_text'])

Limitations

As almost all other language models trained on large text datasets scraped from the web, the Tucano2-qwen-0.5B-Base shows behavior that does not make it an out-of-the-box solution to many real-world applications, especially those requiring factual, reliable, and nontoxic text generation. Tucano2-qwen-0.5B-Base is subject to the following:

Hallucinations: Tucano2-qwen-0.5B-Base can produce content that can be mistaken as facts, but is misleading or entirely false, i.e., hallucinations.
Biases and Toxicity: Tucano2-qwen-0.5B-Base inherits the social and historical stereotypes from the data used to train it. Given these biases, the model can produce toxic content, i.e., harmful, offensive, or detrimental to individuals, groups, or communities.
Language Limitations: Tucano2-qwen-0.5B-Base is primarily designed to interact with Portuguese. Other languages might challenge its comprehension, leading to potential misinterpretations or errors in response.
Repetition and Verbosity: Tucano2-qwen-0.5B-Base may get stuck on repetition loops (especially if the repetition penalty during generations is set to a meager value) or produce verbose responses unrelated to the prompt it was given.

Hence, even though Tucano2-qwen-0.5B-Base is released under a permissive license, we urge users to perform their own risk analysis before using it for real-world applications.

Evaluations

The table below compares the Tucano2 series against other base models of similar size. We divide our evaluations into two sets:

Easy Set: CALAME, GlobalPIQA, LAMBADA, ARC-Challenge, HellaSwag
Hard Set: ENEM, BLUEX, OAB Exams, BELEBELE, MMLU

The NPM (Normalized Performance Metric) provides a balanced view of model performance across tasks, accounting for each task's inherent difficulty by normalizing its evaluation score relative to its random baseline.

	Total Avg.	Easy Set (NPM)	Hard Set (NPM)
Tucano2-qwen-3.7B-Base	59.21	57.41	61
Qwen2.5-7B	57.97	54.12	61.83
Qwen3-4B-Base	57.86	52.52	63.2
SmolLM3-3B-Base	50.25	54.06	46.44
Qwen2.5-3B	50.16	47.69	52.62
Tucano2-qwen-1.5B-Base	47.9	47.97	47.82
Curio-edu-7b	45.66	57.46	33.87
Qwen3-1.7B-Base	44.48	40.94	48.03
Curio-7b	42.79	58.97	26.6
Llama-3.2-3B	40.5	43.79	37.21
granite-3.3-2b-base	39.97	45.31	34.63
Tucano2-qwen-0.5B-Base	35.36	39.93	30.79
Qwen3-0.6B-Base	29.4	26.41	32.38
Llama-2-7b-hf	29.36	42.69	16.03
Tucano2-0.6B-Base	20.64	40.28	0.99
Qwen2.5-0.5B	19.89	18.7	21.09
Curio-1.1b	19.23	39.16	-0.69
Tucano-2b4	17.88	33.55	2.2
Curio-edu-1b1	17.72	34.77	0.67
Llama-3.2-1B	16.57	28.32	4.83
Tucano-1b1	15.44	29.12	1.76
Tucano-630m	14.9	26.99	2.8
Carvalho_pt-gl-1.3B	12.54	26.75	-1.66
TeenyTinyLlama-460m	11.18	19.65	2.72
Tucano-160m	8.78	19.12	-1.56
TeenyTinyLlama-160m	7.72	15.75	-0.31
GlorIA-1.3B	5.93	27.27	-15.42

Evaluation Suite

Benchmark	n-shot	Type	Baseline	Metric
Easy Set
CALAME	5-shot	Completion	0	`acc`
GlobalPIQA	5-shot	Completion	50	`acc_norm`
LAMBADA	5-shot	Completion	0	`acc`
ARC-Challenge	5-shot	MC-Q&A	25	`acc_norm`
HellaSwag	5-shot	Completion	25	`acc_norm`
Hard Set
ENEM	3-shot	MC-Q&A	20	`acc`
BLUEX	3-shot	MC-Q&A	22.5	`acc`
OAB Exams	3-shot	MC-Q&A	25	`acc`
BELEBELE	5-shot	MC-Q&A	25	`acc_norm`
MMLU	5-shot	MC-Q&A	25	`acc`

Individual Benchmarks

	BLUEX	ENEM	OAB	ARC Challenge	BELEBELE	CALAME	Global PIQA	HellaSwag	LAMBADA	MMLU
Tucano2-qwen-3.7B-Base	66.2	77.54	58.45	57.78	83.67	61.08	83	65.32	62.53	65.4
Qwen2.5-7B	65.92	75.02	55.03	54.19	89.67	58.96	78	67.92	59.52	68.55
Qwen3-4B-Base	69.96	77.61	55.58	54.53	87.89	57.95	77	63.19	60.37	68.59
SmolLM3-3B-Base	54.52	61.37	45.51	51.37	77.67	59.15	81	65.57	59.89	56.19
Qwen2.5-3B	58.28	67.32	50.34	45.21	83.22	58.38	75	59.44	57.17	59.79
Tucano2-qwen-1.5B-Base	55.91	68.72	48.29	48.21	74	59.06	77	56.25	54.2	54.04
Curio-edu-7b	47.15	58.64	43.78	50.94	53	60.79	86	66.48	64.62	45.14
Qwen3-1.7B-Base	57.16	65.22	45.79	47.18	77.89	53.56	67	52.55	50.81	55.49
Curio-7b	43.39	50.59	39.68	48.03	45.33	63.44	89	67.58	65.94	40.83
Llama-3.2-3B	50.35	53.04	39.45	41.11	68.89	54.48	69	59.14	59.48	48.28
granite-3.3-2b-base	45.34	54.02	39.54	41.37	65.67	58.77	70	60.81	58.22	45.63
Tucano2-qwen-0.5B-Base	46.87	55.14	40.36	37.44	53.89	58.67	74	48.43	45.14	39.68
Qwen3-0.6B-Base	42.98	49.48	40.46	36.92	65	45.95	54	40.33	41.78	43.54
Llama-2-7b-hf	31.29	31.77	35.49	42.14	41.44	54.53	67	56.76	59.73	38.64
Tucano2-0.6B-Base	21.14	23.58	23.28	37.01	26.22	57.61	79	47.74	39.45	27.18
Qwen2.5-0.5B	32.55	38.91	35.9	28.46	49.56	44.89	44	37.7	39.08	41.17
Curio-1.1b	21.56	21.06	23.1	30.43	22.89	59.25	75	49.45	46.69	26.35
Tucano-2b4	25.45	21.62	26.74	30.43	25.89	50.34	73	48.85	32.39	26.24
Curio-edu-1b1	23.5	19.87	25.01	32.22	26.22	54.91	69	46.3	42.93	25.43
Llama-3.2-1B	24.06	23.93	26.06	31.71	33.33	50	55	45.27	45.6	28.51
Tucano-1b1	25.45	21.55	26.38	30.09	25.67	48.94	68	44.1	28.43	25.26
Tucano-630m	26.7	21.69	26.92	28.72	27.33	47.3	68	40.37	26.2	25.6
Carvalho_pt-gl-1.3B	19.33	18.12	22.32	27.01	26.44	53.42	63	38.53	33.59	24.82
TeenyTinyLlama-460m	25.87	20.15	27.02	27.35	28.11	42.49	59	34.81	21.56	26.65
Tucano-160m	24.76	20.57	17.22	25.56	23.44	43.59	59	33.73	21.64	25.77
TeenyTinyLlama-160m	22.53	18.89	22.32	24.02	26.78	39.79	58	29.89	17.74	25.74
GlorIA-1.3B	4.31	2.52	4.69	26.41	22.78	54.67	64	36.35	36.68	23.69

Performance and Compute

Below, we display the performance of Tucano2-qwen-0.5B-Base across all benchmarks in our evaluation suite. Tucano2-qwen-0.5B-Base is compared with Qwen3-0.6B-Base, the base model from which they were continually pretrained. The percentage variation in performance is displayed in terms of the difference in evaluation scores between the Base and the Continually Pretrained model.

All individual benchmark scores and their evolution across training time can be found in the .plots folder.

Before and After Continual Pretraining

This plot compares the compute requirements (measured as C = 6 * N * D, where N is the number of parameters and D is the number of tokens processed) against the performance of each model (measured by the NPM score).

Performance and Compute Details

	Parameters (B)	Pretraining Tokens (B)	Continual Pretraining Tokens (B)	Total Tokens (B)	Pretraining Compute (FLOPs)	Continual Pretraining Compute (FLOPs)	Total Compute (FLOPs)	NPM Score
Tucano2-qwen-3.7B-Base	3.7	36000	50	36050	8.64e+23	1.11e+21	8.65e+23	59.2
Qwen2.5-7B	7	18000	-	18000	7.56e+23	-	7.56e+23	57.97
Qwen3-4B-Base	4	36000	-	36000	8.64e+23	-	8.64e+23	57.86
SmolLM3-3B-Base	3	11200	-	11200	2.02e+23	-	2.02e+23	50.25
Qwen2.5-3B	3	18000	-	18000	3.24e+23	-	3.24e+23	50.15
Tucano2-qwen-1.5B-Base	1.5	36000	100	36100	3.67e+23	9e+20	3.68e+23	47.89
Curio-edu-7b	7	2000	20	2020	8.4e+22	8.4e+20	8.48e+22	45.66
Qwen3-1.7B-Base	1.7	36000	-	36000	3.67e+23	-	3.67e+23	44.48
Curio-7b	7	2000	150	2150	8.4e+22	6.3e+21	9.03e+22	42.78
Llama-3.2-3B	3	9000	-	9000	1.62e+23	-	1.62e+23	40.5
granite-3.3-2b-base	2	12000	-	12000	1.44e+23	-	1.44e+23	39.96
Tucano2-qwen-0.5B-Base	0.5	36000	50	36050	1.3e+23	1.5e+20	1.3e+23	35.35
Qwen3-0.6B-Base	0.6	36000	-	36000	1.3e+23	-	1.3e+23	29.39
Llama-2-7b-hf	7	2000	-	2000	8.4e+22	-	8.4e+22	29.36
Tucano2-0.6B-Base	0.6	408	-	408	1.47e+21	-	1.47e+21	20.63
Qwen2.5-0.5B	0.5	18000	-	18000	5.4e+22	-	5.4e+22	19.89
Curio-1.1b	1.1	1000	150	1150	6.6e+21	9.9e+20	7.59e+21	19.23
Tucano-2b4	2.4	515	-	515	7.42e+21	-	7.42e+21	17.87
Curio-edu-1b1	1.1	1000	20	1020	6.6e+21	1.32e+20	6.73e+21	17.72
Llama-3.2-1B	1	9000	-	9000	5.4e+22	-	5.4e+22	16.57
Tucano-1b1	1.1	250	-	250	1.65e+21	-	1.65e+21	15.44
Tucano-630m	0.63	211	-	211	7.98e+20	-	7.98e+20	14.89
Carvalho_pt-gl-1.3B	1.3	26	5	31	2.03e+20	3.9e+19	2.42e+20	12.54
TeenyTinyLlama-460m	0.46	6.2	-	6.2	1.71e+19	-	1.71e+19	11.18
Tucano-160m	0.16	169	-	169	1.62e+20	-	1.62e+20	8.78
TeenyTinyLlama-160m	0.16	6.2	-	6.2	5.95e+18	-	5.95e+18	7.71
GlorIA-1.3B	1.3	35	-	35	2.73e+20	-	2.73e+20	5.92

Cite as 🤗

latex

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

Aknowlegments

Polyglot is a project funded by the Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the State of North Rhine-Westphalia (MWK) as part of TRA Sustainable Futures (University of Bonn) and the Excellence Strategy of the federal and state governments.

We also gratefully acknowledge the granted access to the Marvin cluster hosted by University of Bonn along with the support provided by its High Performance Computing & Analytics Lab.

License

Tucano2-qwen-0.5B-Base is licensed under the Apache License, Version 2.0. For more details, see the LICENSE file.