Sabiá 3

Sabiá 3 is the flagship model from Maritaca AI, Brazil's dedicated LLM lab. It is specialized for Brazilian Portuguese: the team reports performance between GPT-4o and Claude 3.5 Sonnet on Portuguese tasks while being far smaller and cheaper to serve — roughly a third to a quarter of the per-token cost of frontier models. The model supports vision/multimodal input and a 128K context window, and Maritaca publishes the Capitu benchmark (Portuguese instruction-following with morphological constraints) for independent comparison.

Worth knowing: Sabiá is API-only — no open weights — so it is a vendor dependency, not a self-hostable asset. The 3.0 generation was discontinued in late 2025 in favor of Sabiá 3.1 and the newer Sabiá-4 family, so new integrations should target the current generation. Pricing is in reais (R$5/1M input tokens for sabia-3.1), which is cheap by US standards but worth converting before budgeting.