IplanRIO released Rio 3.5 Open 397B on June 13, billing it as a government-built frontier AI model with benchmark scores topping established models including Qwen 3.7 Plus. Days after release, AI company Nex published a mathematical proof showing the model is a direct 0.6 Nex / 0.4 Qwen weight merge, with collinearity measurements of 0.993 across all 60 layers and a stable mixing ratio of α ≈ 0.571. IplanRIO subsequently updated the model card to credit Nex, removed the benchmark claims, and attributed the issue to an "incorrect upload" of a base merged version instead of a final distilled model. The dispute centers on attribution standards in open-source AI development, where building on existing open-weight models is common practice but requires explicit credit to all source models under licensing terms like Apache 2.0 and MIT.
IplanRIO Releases Rio 3.5 Model on June 13
Rio de Janeiro's IplanRIO released Rio 3.5 on June 13. The city's IT agency described it as a frontier-class model with 397 billion parameters and a permissive open-source license, developed by the municipal government. The release coincided with Brazil's World Cup opener, and comments about the model spread rapidly from Brazil to international audiences.
The original model card described Rio 3.5 as a post-train of Qwen 3.5 397B, Alibaba's open-base model, with a reasoning layer called SwiReasoning added on top. The reported development cost was R$500,000, approximately $100,000 USD. The architecture uses Mixture-of-Experts, activating around 17 billion of the 397 billion parameters per token. The model supports vision and text, handles over a dozen languages, and ships under an MIT license.
SwiReasoning is a training-free inference framework that switches between two modes. When the model is confident about a next word—low entropy in the probability distribution—it reasons in plain language. When uncertain, it shifts to latent reasoning in hidden internal states without emitting tokens.
The self-reported benchmark scores included Terminal-Bench 2.1 at 70.8%, edging out Qwen 3.7 Plus at 70.3% and DeepSeek v4 Pro at 67.9%. On IMOAnswerBench, Rio 3.5 scored 89.5%. On HLE—Humanity's Last Exam—Rio 3.5 landed at 36.5%, ahead of Qwen 3.7 Plus's 34.7%. Mayor of Rio de Janeiro Eduardo Cavaliere tweeted about the release, stating "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."
Nex Publishes Mathematical Proof of Weight Merge
Nex-AGI, a Shanghai-based open-source AI alliance, posted on X days after the release. The analysis stated: "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." Nex analyzed the weights and reported the formula: Rio 3.5 ≈ 0.6 × Nex N2 Pro + 0.4 × Qwen 3.5. A verification script and a full GitHub report followed.
The evidence included behavioral and mathematical components. Nex stripped the hardcoded "You are Rio" system prompt from the deployed model and sent it 120 identity questions. Without the prompt, Nex reports the model called itself "Nex, from Nex-AGI" 79.2% of the time and "Rio" 0% of the time. The model recited Nex's specific backstory verbatim, mentioning the "Shanghai Innovation Institute" and "a large-model ecosystem alliance."
Mathematically, Nex measured collinearity across all 60 layers. The result came back at 0.993. The mixing ratio held at α ≈ 0.571, stable to three decimal places. Nex stated: "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. There is no innocent explanation."
Nex N2 Pro, released 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, Nex sits at 1,585 against Rio's 1,533.
IplanRIO Updates Model Card and Credits Nex
IplanRIO updated the Hugging Face model card. The benchmark table was removed and the attribution changed. The updated Readme states: "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. 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 from IplanRIO has been released. Nex is now credited in the model card. The "incorrect upload" explanation claims the intended release was a distilled version of the merged base, not the raw merge itself. On-policy distillation involves a stronger teacher model generating outputs while the student trains on those outputs and generates its own.
IplanRIO stated it is working to upload the corrected, distilled model with full attribution in place.
Community Debate on Attribution Standards
Model merging is legal under the licenses involved. Nex N2 Pro is Apache 2.0, permitting use, modification, and redistribution with credit. Qwen 3.5 is openly licensed. The issue centered on presenting the output as independently developed work without naming all source models.
Tech commentator Rafael Quintanilha noted that since Nex N2 Pro is built on Qwen, the team may have credited the underlying architecture and left it there. He pointed out the model went viral during a World Cup match, "not necessarily 'ready for public consumption.'" Developer Lucas Montano stated that "merging two ~400B-class models and then applying policy distillation isn't trivial" while acknowledging both a technical error and a communication failure.
AI researcher Diego Ambrosio noted the original launch described Rio 3.5 as the result of "autonomous post-training and proprietary fine-tuning"—framing that implied original research, not a merge.
Nex wrote on X: "We are flattered that the City of Rio used our work to achieve SOTA performance. But in the open-source world, attribution matters."
FAQ
What did IplanRIO release on June 13?
IplanRIO released Rio 3.5 Open 397B on June 13, described as a government-built frontier AI model with 397 billion parameters, Mixture-of-Experts architecture, and benchmark scores including 70.8% on Terminal-Bench 2.1, 89.5% on IMOAnswerBench, and 36.5% on HLE. The model was released under an MIT license with a reported development cost of R$500,000.
What did Nex's mathematical analysis show about Rio 3.5?
Nex published a mathematical proof showing Rio 3.5 is a direct weight merge with the formula: Rio 3.5 ≈ 0.6 × Nex N2 Pro + 0.4 × Qwen 3.5. The analysis measured collinearity at 0.993 across all 60 layers with a stable mixing ratio of α ≈ 0.571. Identity tests showed the model self-identified as "Nex, from Nex-AGI" 79.2% of the time and "Rio" 0% of the time when the hardcoded system prompt was removed.
How did IplanRIO respond to Nex's findings?
IplanRIO updated the Hugging Face model card to credit Nex, removed the benchmark claims, and stated: "We detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model." The updated card describes the model as "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." IplanRIO stated it is working to upload the corrected distilled model with full attribution.
