Rio de Janeiro's AI model, Rio 3.5, faces scrutiny over its true origins after Nex-AGI reveals it is largely based on their own Nex N2 Pro.
Overview of Rio 3.5 Launch
IplanRIO, the IT agency for Rio de Janeiro, introduced the Rio 3.5 AI model on June 13, boasting 397 billion parameters and an open-source license. The launch coincided with Brazil's World Cup opener, generating considerable interest on social media. Initially, the model was promoted as a government-developed AI that outperformed leading models, including Qwen 3.7 Plus.
However, excitement quickly turned to skepticism regarding the model's origins. The initial description claimed that Rio 3.5 was a refined version of Alibaba's Qwen 3.5, enhanced with a new reasoning layer called SwiReasoning. The development cost was reported at R$500,000 (around $100,000 USD), significantly lower than similar AI solutions on the market.
Performance Claims and Initial Metrics
The architecture of Rio 3.5 utilized a Mixture-of-Experts design, activating only a fraction of its parameters during processing, thereby reducing inference costs. It supports both text and visual inputs and operates in multiple languages under an MIT license. The model's claims included impressive performance metrics, such as scoring 70.8% on the Terminal-Bench 2.1, surpassing Qwen 3.7 Plus and DeepSeek v4 Pro.
Rio 3.5 also excelled in other assessments, including achieving 89.5% on a math olympiad test and 36.5% on a multi-domain test, further fueling the narrative of a municipal government achieving significant advancements in AI.
The Nex Revelation
Shortly after the launch, Nex-AGI, a Shanghai-based AI firm, revealed that Rio 3.5 was fundamentally a blend of their own Nex N2 Pro model and Qwen 3.5. According to Nex, an analysis of the model weights confirmed that Rio 3.5 was approximately 60% Nex N2 Pro and 40% Qwen 3.5. They highlighted that the model's identity was misrepresented, as it frequently identified itself as Nex rather than Rio.
Nex's findings also suggested that Rio 3.5's performance metrics were less favorable when compared to Nex N2 Pro, raising questions about the integrity of IplanRIO's claims. They contended that the model's performance should logically be lower than Nex N2 Pro if it was primarily composed of it.
Response from IplanRIO
In light of these revelations, IplanRIO amended the model card, removing previous benchmark claims and attributing the model's development to both Nex and Qwen. They attributed the confusion to an "incorrect upload" of the model, clarifying that the intended release was a distilled version rather than a straight merge.
The agency acknowledged the mistakes made in communication and technical execution, stating they were sorry for any confusion caused. While the merging of models is legal under their respective licenses, the controversy centers around the lack of transparency regarding the sources utilized in developing Rio 3.5.
Community Reactions and Future Steps
The incident has sparked discussions within the tech community regarding the importance of transparency and attribution in open-source projects. While IplanRIO is working on releasing a corrected version of the model, opinions on the implications of this incident vary widely. Some see it as a technical oversight, while others criticize the misrepresentation of the model's origins.
Nex has chosen not to pursue legal action but emphasized the need for proper attribution in the open-source environment. As IplanRIO prepares to release the revised model, it remains to be seen whether the adjustments will affect its performance metrics or clarify its developmental lineage.
