Drummond: why we trained an open reasoning model in Portuguese

Drummond-1b1-Instruct is an open language model in Brazilian Portuguese, with 1.1 billion parameters, trained to write out its reasoning before the answer. It is a CORRE research project: small, published with code and data, and still far from what you would use in production.

What it is

The repository on GitHub describes it as the first large language model with 100% open reasoning in Brazil, and the model is on Hugging Face. On the CORRE website, it appears in the R&D area, next to Open CORRE, with this description: an open AI model, in Portuguese, for anyone to use and adapt.

How it was trained

Source: GitHub repository and model card on Hugging Face.

The model was trained to produce a chain of reasoning before the final answer, using the special tokens of the ChatML format. A small base model receives a small set of reasoning examples in Portuguese, and the question is what changes in its behavior.

Why an open model, in Portuguese

Three practical reasons make a model like this interesting for people who work in social impact:

1. Language. Most open models are trained mainly in English, and Brazilian Portuguese, with its own text genres, needs specific attention.
2. Auditability. With open code, data and weights, you can see how it was made and what was used, which is not the case with a closed service.
3. Adaptation. A small, open model can be fine-tuned to a specific vocabulary, such as that of a program or a public policy, without asking anyone's permission.

This is an argument for open development, and not a promise about Drummond in particular.

What the tests show

The repository publishes results on the Portuguese evaluation with the LM-Evaluation-Harness, with a context of 2,048 tokens:

The numbers are low. ENEM, for example, has five choices per question, and 20.71% correct is close to what a random guess would give. A model with 1.1 billion parameters and a thousand fine-tuning examples should not be read as a reliable assistant. The model card on Hugging Face has no evaluation numbers and no limitations section. Anyone who uses it should run their own tests.

What it is not

• It is not a model for production.
• It is not an agent: it does not decide or act on its own. In CORRE projects, AI is a block with a defined function, such as transcribing, classifying or summarizing, and the team reviews what it writes.

How to test it

The model loads with the Transformers library, using AutoModelForCausalLM and trust_remote_code=True, and the repository includes an inference example and a Colab notebook. If you study it, adapt it or find a problem, the GitHub repository is the place to tell us. See also Open CORRE, with free social technology resources.