IplanRio originally released its initial lineup of six open-source artificial intelligence models last April during the announcement of Sandbox.Rio, an experimental regulatory environment program promoted by the Municipal Secretariat of Economic Development of the City of Rio de Janeiro.
Building on that momentum, IplanRio announced the reveal of its latest and most powerful LLM this week: the Rio 3.5 Open 397B. This new model beat benchmark results against major international open-source projects such as Alibaba Qwen, DeepSeek V4, and Kim-k2, showing significantly improved speed and overall performance, according to IplanRio's own benchmarks.
IplanRio is a public technology company managed by the City of Rio de Janeiro. According to João Cabaretta, the CEO of IplanRio, the AI project reportedly cost R$ 500,000 (approximately $100,000 USD)—a price tag estimated to be at least 30 times cheaper than developing an off-the-shelf AI system from scratch.
Development and Models
The lineup, now comprising seven LLMs developed by IplanRio, started in April with six initial models. The benchmark-breaking Rio 3.5 Open 397B is the most recent addition, unveiled this June. These models are already being deployed across various administrative tasks by the City of Rio de Janeiro.
During the presentation of the Rio 3 architecture, Rafael Coelho, the project's Chief Scientist, explained that the team had to propose different architectural approaches to make the models truly exceptional. A key innovation was the use of "concept models," which allow the AI to process information in conceptual spaces (such as audio, images, and videos) rather than being limited to isolated words. This breakthrough made Rio 3 approximately eight times faster than traditional models of a similar size.
The team also successfully implemented a context window capable of ingesting up to 1 billion tokens, granting the model significantly larger memory capacity than competing LLMs.
Although the AI was originally developed around the Chinese open-source model Alibaba Qwen, Cabaretta explained that the current version bears little resemblance to the original architecture.
"Previously, AI development was simply based on adding computing power (like GPUs) to improve the model. However, we started hitting a plateau, and developers had to go back to the drawing board to brainstorm new architectural alternatives and ideas. What we did was take the off-the-shelf Qwen model and modify it. Currently, our model has nothing to do with the original Qwen, though we utilized its baseline structure and training," Cabaretta detailed.
The new model is based on Alibaba’s Qwen 3.5 397B (A17B), which was officially announced and released in February 2026, and utilizes a Mixture-of-Experts (MoE) architecture.
Practical Applications in Rio
The model's development is focused on creating a proprietary LLM to improve public administration in Rio de Janeiro. For instance, the system is currently being used by the city government to monitor security cameras for suspicious activities, scan documents for financial accountability, generate institutional images and videos, and power customer service chatbots for public institutions.
"Today, Rio 3 is just as good as the leading models on the market, but much cheaper. We are building specific applications for the city hall, and it has proven to be up to 30 times cheaper than models like Gemini or ChatGPT. Furthermore, it represents the foundation of a 'world model' because it does not think in words, but in concepts, which are then translated into images, words, and videos. This is what academia is currently trying to achieve, and we are heading in that exact direction. This is our first release and an invitation to kick off the construction of this rocket," summarized Cabaretta during his presentation.
The Open-Source Models Lineup
Here are the specifications for the open-source models:
- Rio 3.5 Open 397B: Launched in June 2026 and trained based on the Qwen 3.5 397B, the repository comprises approximately 807 GB distributed across 97 weight files. It features a Mixture-of-Experts (MoE) architecture with a total of 397 billion parameters and about 17 billion active parameters per token.
- Rio 3.0 Open: The open-source flagship model launched in April. It features 235 billion parameters and delivers performance on par with the best open models currently available.
- Rio 3.0 Open Mini: A more compact version with 44 billion parameters. Despite its reduced size, it achieves performance in mathematical benchmarks equivalent to open models up to 10 times larger. It is highly optimized for mobile development.
- Rio 3.0 Open Nano: The most compact model in the lineup, capable of processing 10,000 queries for just R$ 1 ($0.20 USD).
- Rio 3.0 Search: A model specifically designed for web searching.
- Rio 2.5 Open: A model geared towards creativity, designed to run locally on personal computers.
- Rio 2.5 Open VL: A computer vision model capable of generating videos, handling Q&A, and performing Optical Character Recognition (OCR) and Grounding.
