Rio 3.5 Open 397B – from Rio de Janeiro's city government (huggingface.co)

Rio de Janeiro's municipal IT company, IplanRIO, has officially launched the Rio 3.5 Open 397B, a frontier-class general-purpose AI model that builds on the Qwen 3.5 397B architecture. This new model is significant for the AI/ML community as it delivers enhanced performance across various benchmarks, including coding, mathematics, and multilingual capabilities. With 397 billion total parameters and a pioneering SwiReasoning framework, Rio 3.5 Open 397B offers a unique approach to reasoning, switching between explicit and latent modes based on confidence signals. This innovation not only improves accuracy but also optimizes token efficiency, a critical factor in deploying AI at scale. Key technical features include a massive 1 million-token context window and a dynamic reasoning mechanism that enables the model to explore multiple pathways simultaneously during low-confidence scenarios, thus maximizing efficiency. Post-training results showcase substantial improvements over its base model in various benchmarks, achieving notable gains in areas such as software engineering and multilingual reasoning. Furthermore, the model is released under an MIT license, making it accessible for both commercial and research applications, potentially democratizing advanced AI capabilities across languages and disciplines.