Maritaca's Blog
Meet Data Ocean
The platform that connects the Sabiá to dozens of Brazilian databases to answer with real, up-to-date, sourced numbers, in the chat and the API.
Integrated tools in the Sabiá API
The tools the Sabiá already uses in the chat (web search, code execution and Data Ocean) are now available in the API. You activate the ones you need on each request.
Inference 100% in Brazil
Inference 100% on national territory, previously available only via contract and now open to everyone: just call the model with the br-sp suffix.
Sabiá 4 Thinking
The reasoning model in the Sabiá family: frontier-level quality in Portuguese at the lowest cost in its class, with major gains over Sabiá 4 in tool use, legal tasks and response quality.
Why Brazil needs to train its own AIs from scratch, and a plan to make it happen
Why Brazil needs to master the LLM stack before the window closes, and how to organize a competitive three-phase funding program to make it happen.
Scaling LLM training at Maritaca AI
How we build each new generation of the Sabiá family: a benchmark-centric development cycle, a scale pyramid that goes from hundreds of small experiments up to the final training run, and the architecture, data and infrastructure choices that sustain this pace.
Persuasion in LLMs: how far can language models influence opinions
Study on the capacity of language models to persuade users, with controlled experiments and risk analysis.
LLM Bias Bench: measuring opinion bias in language models
We introduce a benchmark for measuring ideological bias in LLMs, with comparative analysis of commercial and open models in Portuguese.
Sabiá-4
We're introducing our new generation of models with Sabiazinho-4 and Sabiá-4, with improvements in the legal domain, long context use, instruction following and agent capabilities.
Sabiazinho-4
We're introducing our new model focused on speed and low cost: Sabiazinho-4, with improvements in legal domain, long context, instruction following and agent capabilities.
