Scandal around Rio 3.5: Municipal AI model turns out to be a "mashup" of two open-source solutions
The municipal IT company of Rio de Janeiro, IplanRIO, presented the Rio 3.5 Open 397B model, claiming it surpasses DeepSeek V4 Pro and Qwen 3.7 Plus on several benchmarks. However, within a day, the AI model developer Nex publicly stated that the "cutting-edge" development is nothing more than a direct merge of the weights of Nex-N2-Pro and Qwen3.5-397B-A17B.
The scandal erupted after IplanRIO uploaded the model to Hugging Face under the MIT license on June 13, 2026. The initial description claimed that Rio 3.5 is a refined version of Qwen3.5-397B-A17B with a Mixture-of-Experts (MoE) architecture, 397 billion parameters, and a context window of 1.01 million tokens. The Mayor of Rio de Janeiro, Eduardo Cavalieri, had already announced on social media platform X that "the open AI model, trained with public funds, has surpassed all others."
Evidence from Nex
However, on June 14, Nex published a statement in the Nex-N2 repository on GitHub, providing irrefutable evidence. The company stated that Rio 3.5's weights consist of 60% Nex-N2-Pro and 40% Qwen3.5-397B-A17B. Moreover, after removing the system prompt "You are Rio," the model identified itself as "Nex, from Nex-AGI" in 79% of responses. Analysis showed that each tensor of Rio's weights replicates the 0.6/0.4 ratio across all 60 layers.
"There is no innocent explanation for this," said Nex in a statement. "Rio 3.5 is our open-source model in a different wrapper."
Issue with Benchmarks
Additional questions were raised by the test results themselves. It turned out that the original Nex-N2-Pro model scores 75.3% on Terminal-Bench 2.1, while Rio 3.5 scores only 70.8%. On GDPval, the gap was 1585 points versus 1533 points. If Rio is indeed a mixture, its weaker performance seems logical. After the exposure, IplanRIO removed the benchmarks from the main description.
Reaction from IplanRIO
Under the pressure of facts, IplanRIO updated the model card on Hugging Face. The new version of the README states that Rio 3.5 Open 397B was built through a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by distillation from a stronger model. The team apologized: "We regret the confusion" and promised to upload a corrected version.
Analysis of the Situation
It is important to understand: using open models is not a violation. Both models, Nex-N2-Pro and Qwen3.5-397B-A17B, are distributed under open licenses. The problem arose because IplanRIO presented a simple compilation as the result of its own deep research work. In the open-source community, transparency is a key principle. Weight merging and distillation are common practices, but developers must honestly indicate which components they are using and where they come from.
My expert opinion: This incident is a vivid example of how the pursuit of PR effect can undermine trust in a project. The municipality of Rio likely wanted to showcase achievements in AI but chose the wrong path. For the community, this is a lesson: "breakthrough" models trained with public funds should undergo independent auditing before publication, not after.