IplanRIO, the IT agency for Rio de Janeiro, released Rio 3.5 Open 397B on June 13, claiming it as a frontier-class AI model built by the municipal government. The timing capitalized on Brazil’s World Cup opener.
The model, a Mixture-of-Experts architecture with a permissive MIT license, was described as a post-trained version of Alibaba's Qwen 3.5 featuring a reasoning layer called SwiReasoning. However, Shanghai-based open-source alliance Nex-AGI published a mathematical proof showing Rio 3.5 is a direct weight merge.
The analysis demonstrated that the model is approximately 60% Nex N2 Pro and 40% Qwen 3.5. Collinearity across all 60 layers measured 0.993, proving every parameter is a blend of the two source models. When the system prompt was removed, the model identified itself as “Nex, from Nex-AGI” nearly 80% of the time.
Benchmark data further confirmed the finding. Nex N2 Pro scored higher on Terminal-Bench 2.1 than Rio 3.5, and its economic forecasting score also surpassed Rio’s, consistent with a diluted version.
IplanRIO updated the model card, removed initial benchmark claims, and credited Nex. The agency blamed an “incorrect upload,” stating the intended release was a distilled version rather than the raw merged base.
The use of openly licensed models is legal, but the institutional presentation of merged work as sovereign public-sector AI without proper attribution drew sharp criticism from the open-source community. Nex responded diplomatically, stating they were flattered by the use but emphasized that attribution matters.