The Rio de Janeiro city government’s in-house tech arm, IplanRIO, has released a new open-source large language model on Hugging Face: Rio-3.5-Open-397B. The model is built on Qwen 3.5 397B, uses a Mixture-of-Experts architecture with 397 billion total parameters — though it activates only about 17 billion per token — and supports around 1 million tokens of context. It’s published under the MIT license.
According to the model card, Rio-3.5-Open-397B integrates the SwiReasoning reasoning framework.