Rio's Record-Breaking AI Model Was Hailed as a DeepSeek Beater. Then Someone Checked the Weights Brazil's Rio de Janeiro government launched an AI model called Rio 3.5 claiming it outscored DeepSeek and Qwen, but a weight analysis found it was roughly 60% Nex N2 Pro blended with the rest being Qwen. On June 13, IplanRIO, the IT agency of Rio de Janeiro, put out an AI model called Rio 3.5 and described it as a frontier-class system. By the agency's account it carried 397 billion parameters, came with a permissive open-source license, and had been built by the municipal government of a city in the Global South. The timing could not have been better. Brazil was playing its World Cup opener, social media was already buzzing, and chatter about the model spread quickly out of Brazil and around the world. But the praise arrived alongside an equally fast question: who actually created the thing? The Original Pitch: Cheap and Open The first model card described Rio 3.5 as a post-train of Qwen 3.5 397B, Alibaba's open-base model, with a new reasoning layer called SwiReasoning placed on top. Development cost was reported at R$500,000, a figure Rio never confirmed, which works out to nearly $100,000 USD, roughly 30 times cheaper than comparable off-the-shelf AI systems. Its design is Mixture-of-Experts, meaning only around 17 billion of the 397 billion parameters actually fire on any given token. That keeps inference far cheaper than the headline size implies. The model handles both vision and text, works across more than a dozen languages, and ships under a fully open MIT license. What SwiReasoning Promised Technically, SwiReasoning was the heart of the claim. It is a training-free inference framework that switches automatically between two modes. When the model is confident about the next word, that is, when entropy in the probability distribution is low, it reasons in plain language. When it is uncertain, it shifts into latent reasoning, thinking in hidden internal states without emitting any tokens. IplanRIO said Rio 3.5 had been trained specifically to take advantage of this, and that the payoff showed up in the benchmark numbers. The Numbers That Made Headlines The self-reported results were striking. On Terminal-Bench 2.1, which measures autonomous terminal command execution and is scored as the percentage of tasks passed, Rio 3.5 came in at 70.8%, edging past Qwen 3.7 Plus at 70.3% and the powerful DeepSeek v4 Pro at 67.9%. On IMOAnswerBench, a math olympiad benchmark scored as percentage correct, Rio 3.5 reached 89.5%. On HLE, Humanity's Last Exam, a near-unsolvable multi-domain expert battery scored as a percentage, Rio 3.5 landed at 36.5%, ahead of Qwen 3.7 Plus at 34.7%. A municipal government topping the most important flagship models on the most meaningful quality benchmarks was the story that travelled, especially once the Mayor of Rio de Janeiro tweeted about it. "An open AI model trained in Rio and publicly funded over the last year by [the Municipality of Rio] has just surpassed all other models," Eduardo Cavaliere wrote. "Today, the world is talking about an open AI model trained in Rio." The "Trained in Rio" Claim Cracks That phrase, "trained in Rio," turned out to be not entirely accurate. Nex-AGI, a Shanghai-based open-source AI alliance, posted on X days after the release. The opening line set the tone: "The Rio 3.5 model broke the internet this week. The plot twist? It's essentially our open-source model, Nex N2 Pro, wearing a different hat." They had analyzed the weights, and the math was exact: Rio 3.5 ≈ 0.6 × Nex N2 Pro + 0.4 × Qwen 3.5. A verification script and a full GitHub report followed. Two Lines of Evidence The first was behavioral. Nex stripped the hardcoded "You are Rio" system prompt out of the deployed model and asked it 120 identity questions. Without the mask, Nex reports the model called itself "Nex, from Nex-AGI" 79.2% of the time. It called itself "Rio" exactly 0% of the time. Nex also says the model recited the company's specific backstory word for word, naming the "Shanghai Innovation Institute" and "a large-model ecosystem alliance." That is Nex's own training data, surfacing inside someone else's model. The second was mathematical. In a genuine weight merge, every parameter in the new model sits on a straight line between the two source models. Nex measured this collinearity across all 60 layers and got 0.993. Two unrelated models in the same parameter space score near zero by chance, so hitting 0.993 on every single layer is no coincidence. The mixing ratio held steady at α ≈ 0.571, stable to three decimal places. In plain terms, it was close to 60% Nex, with the remainder being the base Qwen model. "Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen, across all 60 layers and every component of the network," Nex wrote. "There is no innocent explanation." The numbers carried a quieter story too. Nex N2 Pro, released just days before Rio 3.5, scores 75.3% on Terminal-Bench 2.1, higher than Rio's 70.8%. On GDPval, an economic forecasting benchmark scored as an Elo-style rating, Nex sits at 1,585 against Rio's 1,533. If Rio is 60% Nex, you would expect it to score below Nex on Nex's own benchmarks, and it does. Rio's Defense: An "Incorrect Upload" IplanRIO updated the Hugging Face model card. The benchmark table came down and the attribution changed. "The model is built via a merge of nex-agi/Nex-N2-Pro and Qwen/Qwen3.5-397B-A17B, preceded by On-Policy Distillation from a stronger model," the updated Readme says. "We detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model. We are sorry for the confusion and apologize profusely." No other public statement has come from IplanRIO. Nex is now credited. The "incorrect upload" explanation is the key claim. IplanRIO says the intended release was a distilled version of the merged base, not the raw merge itself. On-policy distillation means a stronger teacher model generates outputs and the student model trains on those while also producing its own. It costs more than a raw merge but still less than training from scratch. If that step really happened, it would represent at least some original work layered on top of the merge. But by IplanRIO's own account, what actually shipped was the merged base with nothing on top. A Divided Community Observers split over what that meant. Tech commentator Rafael Quintanilha offered the charitable read: since Nex N2 Pro is itself built on Qwen, the team may have credited the underlying architecture and left it there. He also pointed out that the model went viral during a World Cup match and was "not necessarily 'ready for public consumption.'" Developer and AI YouTuber Lucas Montano noted that "merging two ~400B-class models and then applying policy distillation isn't trivial," while still acknowledging both a technical error and a communication failure. AI researcher Diego Ambrosio was less generous. The original launch had described Rio 3.5 as the result of "autonomous post-training and proprietary fine-tuning," framing that implied original research rather than a merge. Merging Is Legal. Hiding It Is the Problem Model merging is completely legal. Nex N2 Pro is Apache 2.0, so you can use it, modify it, and redistribute it as long as you credit it. Qwen 3.5 is openly licensed as well. Nobody is going to court. The real problem was presenting the output as independently developed work without naming all the source models. The open-source community has been here before. Earlier this year, Cursor's Composer 2 was found to be built on Moonshot's Kimi K2.5 without disclosure. The backlash was fast and reputational, no lawyers, just screenshots. Building on existing open models is normal. As TrendKia has covered, stacking and merging open weights is practically a subculture of its own. The norm is not "don't build on others' work." The norm is: say what you used. The Institutional Wrapper Made It Louder What turned this into more than a routine attribution miss was the institutional wrapper around it. A pseudonymous developer shipping a frankenmerge under their own name is one thing. A municipal government using it to claim public-sector AI sovereignty, during the World Cup no less, is another. "It was a waste of resources," one Brazilian commentator wrote. Nex declined to turn it into a war. "We are flattered that the City of Rio used our work to achieve SOTA performance," the company wrote on X. "But in the open-source world, attribution matters." IplanRIO is now working to upload the corrected, distilled model with full attribution in place. When that lands, the same checks will run again, and the community will find out whether the distillation actually changed anything, or whether it is still mostly Nex with a different system prompt. What this means for you This story matters most to AI developers, researchers, and anyone following open-source technology. • For developers and researchers: Before trusting a model's inflated benchmark claims, it pays to check its source and license, since a weight analysis can expose what a system is really built from. • For everyday users: Models like Rio 3.5, Nex N2 Pro, and Qwen 3.5 ship under open licenses, so using or merging them is legal as long as the original sources are credited. Questions & Answers 1. What is Rio 3.5 actually built from? According to a weight analysis it is roughly 60% Nex N2 Pro blended with Qwen 3.5, precisely Rio 3.5 ≈ 0.6 × Nex N2 Pro + 0.4 × Qwen 3.5. 2. How did Nex-AGI prove the copy? Nex offered two kinds of evidence: without the mask the model identified itself as Nex 79.2% of the time and as Rio 0% of the time, and collinearity across all 60 layers came back at 0.993, pointing to a genuine merge. 3. Is merging models like this illegal? No. Nex N2 Pro is Apache 2.0 and Qwen 3.5 is openly licensed, so using and merging them is legal as long as you credit them. The real issue was failing to name the sources. 4. How did IplanRIO explain it? It called the situation an "incorrect upload," saying the base merged version was uploaded by mistake instead of the final distilled model, and it is now working to upload the corrected, distilled model with full attribution. https://trendkia.com/en/ai/riyo-ke-sabase-takatavara-ai-modala-ka-sacha-deepseek-ko-harane-ka-dava-asala-me-1090 TrendKia — Har trend, sabse pehle.