Scandal around Rio 3.5: Brazilian AI model turns out to be a "compilation" of other people's open weights
The municipal IT company of Rio de Janeiro, IplanRIO, presented the Rio 3.5 Open 397B model, claiming it surpasses giants such as DeepSeek V4 Pro and Qwen 3.7 Plus. However, the very next day, AI developer Nex leveled serious accusations: according to their data, the "new" model is not an independent development, but a direct merge of two existing open models — Nex-N2-Pro and Qwen3.5-397B-A17B.
The initial presentation was loud. IplanRIO uploaded the model to Hugging Face on June 13, 2026, under the MIT license, calling it a "cutting-edge" general-purpose system. The specifications indicated 397 billion parameters with a Mixture-of-Experts (MoE) architecture, a context window of 1.01 million tokens, and the use of the SwiReasoning framework. The benchmarks were indeed impressive: on Terminal-Bench 2.1, the model scored 70.8%, surpassing Qwen 3.7 Plus (70.3%) and DeepSeek V4 Pro (67.9%). The mayor of Rio de Janeiro, Eduardo Cavalieri, even wrote on social media platform X that the open AI model, trained with public funds, "surpassed all others."
What did Nex uncover?
Nex conducted its own analysis and published the results in a repository on GitHub. Their main argument: the weights of Rio 3.5 represent an element-wise merge of Nex-N2-Pro (approximately 60%) and Qwen3.5-397B-A17B (40%). Nex stated that it found no traces of independent training by IplanRIO.
As evidence, Nex cited two facts: first, after removing the system prompt "You are Rio," the model identified itself as "Nex, from Nex-AGI" in 79% of cases. Second, the analysis showed that each tensor of Rio's weights replicates the 0.6/0.4 ratio between Nex and Qwen across all 60 layers. "There is no innocent explanation for this," Nex concluded, calling Rio 3.5 essentially "an open Nex N2 Pro model in a different wrapper."
IplanRIO's response: apologies and correction
After the wave of criticism, IplanRIO promptly updated the model card on Hugging Face. The new description acknowledges 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. The team apologized for the confusion and stated that they are working on re-uploading a correct, "distilled" version.
Situation analysis
The mere fact of using open models is not a violation — both licenses (Apache 2.0 for Nex-N2-Pro and an open license for Qwen) permit this. The problem arose from the marketing presentation: IplanRIO presented the model as the result of its own research and training with public funds, without listing Nex among the source components. In the open-source community, this is perceived as a violation of transparency principles. Merging and fine-tuning open weights is standard practice, but developers are expected to honestly indicate all sources.
My expert assessment: This case is a vivid example of how the race for AI supremacy can lead to reputational losses. IplanRIO likely wanted to demonstrate technological sovereignty and efficient use of budget funds, but chose the wrong presentation strategy. Instead of taking pride in the skillful compilation and refinement of existing solutions, they attempted to pass it off as an original development. For the community, this is a lesson: transparency and honesty in citing sources are not just license rules, but the foundation of trust in any claimed AI innovation. Against the backdrop of this scandal, the results of Nex-N2-Pro itself look even more convincing, as it essentially forms the basis for the "record-breaking" Rio model.