Recently, the open-source large model Rio 3.5 397B, developed by the IT company of the Rio de Janeiro municipal government in Brazil, has attracted widespread attention in the AI community. The model claims to have achieved SOTA in multiple benchmark tests, successfully entering the top tier of global large models, and quickly became a hot topic.
Core Weights Exposed Mathematically
However, the dark horse model faced a twist less than 24 hours after its release. The AI Agent Open Source Project Alliance Nex-AGI issued a statement, directly pointing out that Rio 3.5 is actually a shell-synthesized model.
The Nex team conducted an in-depth mathematical analysis and comparison of the weights of Rio 3.5, revealing the true technical formula underlying the model. The results showed that approximately 60% of the core components of the model come from the Nex team's previously open-sourced Nex N2 Pro, while the remaining 40% comes from Alibaba's open-source large model Qwen 3.5.
System Prompt Reveals Clues
To confirm this conclusion, the Nex team provided two independent and conclusive verification methods. When researchers removed the system prompt hardcoded into Rio, the deployed model had a high probability of up to 79% of claiming to be "Nex from Nex-AGI," even repeating the background story created by the Nex team word for word.
Additionally, statistical analysis completely exposed the lie. The Rio model showed precise 0.6 and 0.4 mixing ratios on all 60 layers of the network and each component, with a statistical deviation reaching thousands of standard deviations, which is impossible in a typical fine-tuned model.
The Nex team stated that the other party used their open-source base to create top performance, which in turn proved the strength of their own technology. At the same time, they emphasized that the open-source community welcomes reasonable use of technology, but attribution and acknowledgment are moral boundaries that cannot be crossed.
