The open source model Rio 3.5 397B launched by the IT company owned by the Rio de Janeiro Municipal Government in Brazil hit the AI circle yesterday, claiming to have surpassed Qwen 3.7 Plus and other benchmark tests and won SOTA, entering the first echelon in the world.However, in less than 24 hours, Nex-AGI, an AI agent open source project alliance initiated by Shanghai Chuangzhi Institute, issued a statement identifying Rio 3.5 as a shell suture model.
The Nex team conducted a mathematical analysis and comparison of the weight of Rio 3.5 and found that its underlying technical formula is: Rio 3.5≈0.6×Nex N2 Pro + 0.4×Qwen 3.5.
In other words, 60% of the core genes of this dark horse model, which claims to bring a huge breakthrough, come from the Nex team’s previously open source Nex N2 Pro, and 40% come from Alibaba’s Qwen 3.5.
The Nex team provided two independent verification methods. First, after removing Rio's hard-coded "You are Rio" system prompt word, the deployed model has a 79% probability of calling itself "Nex from Nex-AGI" and a 0% probability of calling itself Rio. It will even retell Nex-AGI's specially customized background story verbatim.
Second, every weight tensor in Rio shows the same 0.6/0.4 mixing ratio in all 60 layers and every component of the network. The statistical deviation reaches thousands of standard deviations. Other fine-tuning models cannot be explained by this interpolation method.
The Nex team said that Rio 3.5 uses Nex N2 Pro to piece together SOTA performance, which proves the power of the Nex open source base. It also emphasizes that the open source community has its own operating rules and everyone is welcome to use it, but signature and acknowledgment are insurmountable bottom lines.