Rio 3.5: The scandal over Rio de Janeiro City Hall's "unique" AI model
Rio de Janeiro's municipal IT company IplanRIO has introduced Rio 3.5 Open 397B, claiming it is an open AI model trained with public funds and surpassing DeepSeek V4 Pro and Qwen 3.7 Plus on several benchmarks. However, the very next day, the AI development team Nex stated that the tool is a direct merge of their own model Nex-N2-Pro and Qwen3.5-397B-A17B.
Following these claims, IplanRIO promptly updated the model card on Hugging Face. The new version states that Rio 3.5 was built through a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by distillation from a stronger model.
How Rio 3.5 was presented
IplanRIO uploaded Rio 3.5 Open 397B to Hugging Face on June 13, 2026, under the MIT license. The initial description called the project a "cutting-edge" general-purpose AI system and indicated that the model was fine-tuned based on Qwen3.5-397B-A17B. Specifications included 397 billion parameters, of which 17 billion are activated when processing each token (Mixture-of-Experts architecture), a context window of 1.01 million tokens, and the use of the SwiReasoning framework. The first version of the card presented test results where Rio 3.5 outperformed competitors: on Terminal-Bench 2.1 — 70.8% versus 70.3% for Qwen 3.7 Plus and 67.9% for DeepSeek V4 Pro; in Humanity's Last Exam — 36.5% versus 34.7% for Qwen 3.7 Plus; in IMOAnswerBench — 89.5%.
What Nex stated
On June 14, Nex opened an issue in the Nex-N2 repository on GitHub, stating that the weights of Rio 3.5 appear to be a direct element-wise merge of Nex-N2-Pro and Qwen3.5-397B-A17B. According to Nex's assessment, Rio 3.5 consists of approximately 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. The company found no evidence of independent training by IplanRIO. Two key arguments: after removing the system prompt "You are Rio," the model identified itself as "Nex, from Nex-AGI" in 79% of responses, and every tensor of Rio's weights replicates the 0.6/0.4 ratio between Nex and Qwen across all 60 layers of the model. "There is no innocent explanation for this," Nex's statement reads.
Why the benchmarks raised questions
Nex-N2-Pro shows higher results in its own tests than Rio 3.5: 75.3% on Terminal-Bench 2.1 versus 70.8% for Rio 3.5, and 1585 points on GDPval versus 1533. If Rio is indeed a mixture, its weaker results appear expected. After the card update, these benchmarks were removed from the main description.
How IplanRIO responded
Following the claims, IplanRIO changed the model's README, stating that Rio 3.5 was built through a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, and then underwent distillation. "We regret the confusion and apologize," the updated README says. The team is working on re-uploading the correct model. No separate detailed public comment was available at the time of publication.
What the dispute is about
The use of open models (Nex-N2-Pro under the Apache 2.0 license, Qwen3.5-397B-A17B — an open model) is not a violation in itself. The dispute arose from the presentation: the initial card created the impression of independent development and fine-tuning based on Qwen, but did not mention Nex-N2-Pro as one of the sources. In the open-source community, this is perceived as a transparency issue. Merging open weights and distillation have become common practice, but developers are expected to disclose the original models and the contributions of third-party teams.
My comment: This incident is a vivid example of how transparency suffers in the race for AI leadership. For the crypto community, where verification and proof are the foundation of trust, such an approach is unacceptable. If IplanRIO had honestly disclosed the sources from the start, there would have been no scandal. Instead, we see an attempt to pass off a combination of open models as one's own breakthrough, which undermines trust not only in the project but also in government initiatives in this field.