Sabi\'a-3 Technical Report

TL;DR

This paper introduces Sabiá-3, a large Brazilian Portuguese language model that achieves high performance on various benchmarks at a significantly lower cost, emphasizing the advantages of domain-specific training.

Contribution

The paper presents Sabiá-3 and Sabiazinho-3, new language models trained on a Brazilian corpus, with Sabiá-3 outperforming previous models and matching frontier LLMs at reduced costs.

Findings

Sabiá-3 performs strongly on Portuguese and Brazil-related tasks.

Sabiá-3 matches frontier LLM performance levels.

Cost per token is three to four times lower than comparable models.

Abstract

This report presents Sabi\'a-3, our new flagship language model, and Sabiazinho-3, a more cost-effective sibling. The models were trained on a large brazilian-centric corpus. Evaluations across diverse professional and academic benchmarks show a strong performance on Portuguese and Brazil-related tasks. Sabi\'a-3 shows large improvements in comparison to our previous best of model, Sabia-2 Medium, especially in reasoning-intensive tasks. Notably, Sabi\'a-3's average performance matches frontier LLMs, while it is offered at a three to four times lower cost per token, reinforcing the benefits of domain specialization.