Rio de Janeiro's "in-house" LLM appears to be a merge of existing models

Analysis suggests that the Rio-3.5-Open-397B model released by the Rio de Janeiro city government via IplanRIO is not a model trained in-house, but rather the result of weight merging the existing models Nex-N2-Pro and Qwen3.5-397B-A17B at a 6:4 ratio.

According to the weight analysis, the model's major tensors showed a very high correlation (cos_fit ≈ 0.99) with a linear combination of the two models, a figure that is difficult to explain through simple fine-tuning. In addition, when the model's identity was queried with the system prompt removed, it responded with the identity of Nex-AGI rather than Rio in the majority of cases.

Regarding this, the city government explained that a base merged version was mistakenly uploaded instead of the final distilled model during the upload process, but the community continues to express doubts about whether the model was actually trained in-house.