A Brazilian government-backed AI model hailed as Latin America's open-source breakthrough was revealed to be a weight mixture of two existing models, triggering an apology from its developer within 24 hours.
Back
IplanRIO's Rio 3.5 model exposed as 60% Nex-AGI, 40% Qwen shell
A Brazilian government-backed AI model hailed as Latin America's open-source breakthrough was revealed to be a weight mixture of two existing models, triggering an apology from its developer within 24 hours.
The Rio-3.5-Open-397B model, released June 15 by IplanRIO — the IT arm of Rio de Janeiro's municipal government — was exposed as a roughly 60:40 weight mixture of Nex-AGI's Nex-N2-Pro and Alibaba's Qwen3.5-397B-A17B, rather than an original development. The revelation, published by AI research group Nex-AGI on GitHub, showed that 79% of the model's responses identified as "Nex from Nex-AGI" when its fixed "You are Rio" instruction was removed, with zero responses identifying as Rio.
"We found that across all 60 layers, the weight tensors matched a 60% Nex-N2-Pro and 40% Qwen mixture with a degree of agreement that cannot be explained by typical additional training," Nex-AGI said in its analysis. "No evidence of proprietary training could be found."
IplanRIO had initially described Rio 3.5 Open 397B as a model based on Alibaba's Qwen3.5-397B-A17B with additional training, claiming it surpassed the underlying Qwen in programming and mathematics benchmarks. The model gained rapid attention as a Latin American open-source AI contender before the analysis emerged. In response, IplanRIO said it had performed "on-policy distillation" — a process of mixing models and training on outputs from a more powerful AI — and that the publicly uploaded file was an incomplete pre-distillation version mistakenly uploaded.
The controversy erupted less than 24 hours after the model's Hugging Face debut, when Nex-AGI published a detailed weight analysis showing the model's internal structure was nearly indistinguishable from a linear combination of its two predecessors. Since Nex-N2-Pro is itself based on the Qwen3.5 series, the two models share sufficiently similar architectures to allow weight mixing.
Developing a large-scale language model from scratch requires massive training data and high-performance computing resources, making model merging — a technique that combines learned weights in specific proportions — a common shortcut. IplanRIO's claim of on-policy distillation would represent a legitimate development pathway if verified, but the company has not yet released the post-distillation version it promised.
The episode damages trust in newly hyped open-source models from emerging AI markets. For investors tracking the open-source LLM sector, the incident highlights the gap between claimed capabilities and verified performance — a risk that applies across the rapidly expanding field of government and institutional AI projects. Alibaba's Qwen, already one of the most widely adopted open-source model families globally, now faces the challenge of policing unauthorized commercial reuse of its weights, though the direct financial impact on Alibaba Group Holding Ltd. — which trades at roughly 11 times forward earnings — is likely minimal.
This article is for informational purposes only and does not constitute investment advice.
