Scandal with Rio 3.5: Municipal AI model turned out to be a "mashup" of open-source components
The municipal IT company of Rio de Janeiro, IplanRIO, found itself at the center of a controversial situation after presenting its open-source AI model, Rio 3.5 Open 397B. Hailed as a breakthrough, trained with public funds and surpassing giants like DeepSeek V4 Pro and Qwen 3.7 Plus, the model faced serious accusations of lacking originality within just a day.
The development team at Nex, a reputable player in the AI field, stated that the weights of Rio 3.5 represent a direct element-wise merge of two well-known open-source models: their own Nex-N2-Pro and Qwen3.5-397B-A17B from Alibaba. According to Nex's assessment, the proportion was 60% and 40%, respectively. As evidence, they cited two facts: after removing the system prompt, the model identified itself as "Nex" in 79% of cases, and analysis of the weight tensors showed an identical proportion across all 60 layers.
From bold claims to quiet admission
The initial model card on Hugging Face was full of ambition. It claimed that Rio 3.5 was a "state-of-the-art" AI system with a Mixture-of-Experts (MoE) architecture of 397 billion parameters (of which 17 billion are active per token), a context window of 1.01 million tokens, and its own reasoning framework, SwiReasoning. The mayor of Rio de Janeiro even rushed to announce the "success" on social media, stating that the model "surpassed all others."
However, under pressure from irrefutable evidence from Nex, IplanRIO was forced to update the model description. The current version of the README on Hugging Face explicitly states that Rio 3.5 was built through the merging of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by distillation. The team apologized for the "confusion" and promised to upload a correct, independently trained version.
Situation analysis: Transparency first
The mere fact of using open-source models to create new ones is not a violation. The Apache 2.0 and MIT licenses under which the original components are distributed allow this. The problem lies in the presentation. IplanRIO created the impression that it had carried out extensive independent training from scratch, while in reality, it performed a merge and, likely, minimal fine-tuning. This undermines trust in the project and raises questions about the efficient use of budget funds.
This incident is a vivid example of how, in the modern world of AI, where open-source models become building blocks, transparency and honest attribution are of paramount importance. The race for flashy headlines should not overshadow the fundamental principles of academic and engineering ethics.