AI from Brazil, from the ground up: open-weight language models, civic technology, and the systems that run them.

maracatu.org · Hugging Face

Maracatu is an independent Brazilian effort at the intersection of three themes: foundation models native to Brazilian Portuguese, AI for civic transparency, and the systems software underneath. We open our foundations — the models, the design system, the low-level tooling — and build our applications on top of them. We contribute to the Brazilian AI Plan (PBIA 2024–2028).

Language models trained from scratch in Brazilian Portuguese. Llama-style decoder-only transformer, Apache 2.0 across code and weights. Released so far: Maracatu-20M (17M params, val PP 23.81) and Maracatu-80M (87.8M params, val PP 21.34 on a 3.27M-token holdout — outperforming Tucano-160M with roughly half the parameters).

Design tokens, theme, and shared React primitives that give every Maracatu interface one visual language. Published to npm as @maracatu/estandarte.

A single-host TCP connection dispatcher in Rust. io_uring AcceptMulti hands file descriptors to local workers via SCM_RIGHTS. Round-robin, zero byte copying, no HTTP parsing. ~300 lines, two dependencies, no async runtime — for resource-constrained environments where every syscall counts.

Our flagship: a conversational platform that makes Brazilian public spending legible to any citizen. Ask in plain Portuguese — "how much did deputy X spend in 2025?", "is the company that won this contract clean?", "who voted against this bill?" — and get answers with real data, charts, and links to official sources, drawn from the Câmara, Senado, Portal da Transparência, PNCP, Receita Federal, SICONFI, TSE, TCU and BNDES.

Our projects are named after elements of the Maracatu, an Afro-Brazilian cultural tradition from Pernambuco recognized as intangible cultural heritage by UNESCO. Calunga is the sacred doll carried at the front of the procession; baque is the constant drumbeat; terreiro is the sacred ground where the maracatu gathers; cortejo is the public procession; estandarte is the banner it marches behind. Catraca (turnstile) comes from a different lineage — pure systems engineering — but shares the theme: controlling who passes, and where.

• Star a project that interests you.
• Read each repository's CONTRIBUTING.md for how to send patches.
• Found a security issue? See each repository's SECURITY.md.

Each open project is licensed independently — see its LICENSE file. The maracatu-llm repository (code and weights) is Apache 2.0.