Rio de Janeiro Launches AI Model That Sparks Ownership Controversy

On June 13, 2026, Rio de Janeiro's municipal IT agency IplanRIO unveiled the Rio 3.5 AI model, which it touts as a frontier-class system featuring 397 billion parameters. The initial promotion asserted that Rio 3.5 achieved benchmark scores surpassing those of Alibaba’s DeepSeek. This release coincided with Brazil's World Cup opener, garnering immediate social media traction.

However, Nex, an AI company, contested the originality of Rio 3.5, providing mathematical proof indicating that the model is fundamentally a composite, employing a 0.6 weight from its system and 0.4 from Qwen 3.5, Alibaba's open-source model. Following this revelation, IplanRIO updated the model’s documentation, credited Nex, and retracted earlier benchmarking assertions due to what it labeled an "incorrect upload." IplanRIO originally advertised Rio 3.5 as a post-training enhancement of Qwen 3.5, augmented by an innovative reasoning layer called SwiReasoning. The development reportedly cost R$500,000 (approximately $100,000 USD), a fraction of the expense of similar commercial AI systems. Notably, the model uses a Mixture-of-Experts architecture, allowing only about 17 billion of its parameters to function per token, which arguably lowers operational costs.

Rio 3.5 includes capabilities for processing both visual and textual inputs in over a dozen languages and operates under an open MIT license. The SwiReasoning framework is designed to enhance performance by dynamically switching between direct reasoning in clear language and latent reasoning, where the model internally processes data without directly outputting text.

The model's self-reported scores were particularly notable, claiming a 70.8% success rate on Terminal-Bench 2.1 compared to Qwen 3.7 Plus's 70.3% and DeepSeek v4 Pro at 67.9%. On IMOAnswerBench, an examination benchmark, it achieved an 89.5% success rate.