Rio de Janeiro’s municipal claim of a 397‑billion‑parameter LLM turns out to be a stitched‑together copy. IplanRIO’s “Rio‑3.5‑Open‑397B” package contains weights that are a 0.6 Nex / 0.4 Qwen element‑wise merge of Nex‑AGI’s model with the official Qwen3.5‑397B‑A17B base. No independent training data appears in the file in the repository or on any public platform without any additional fine‑tuning steps.

Removing Rio’s branded “You are Rio” prompt reveals the model’s true identity. When queried, it identifies itself as “Nex, from Nex‑AGI” 79% of the time and never as “Rio.” The system even reproduces IplanRIO’s custom backstory verbatim, indicating that the supposed local training was merely a veneer over the inherited weights across all deployment instances in Brazil and beyond without.

Weight‑level analysis confirms the blend: every tensor in Rio mirrors the 0.6/0.4 mix across all 60 layers, a pattern impossible to achieve through typical fine‑tuning alone. This discovery undermines Rio’s claim of an original, locally trained LLM and highlights the importance of transparent provenance to ensure public trust and prevent misrepresentation in AI policy decisions.