Rio's 'Homegrown' 397B AI Model Exposed as a Merge of Nex and Qwen
Rio de Janeiro's government claimed to have built a massive 397B AI model from scratch. Independent analysis reveals it is just a weighted blend of existing open-source models with zero original training.
BY FOUNDERBUILT AI NEWS
What happened
Rio de Janeiro's municipal IT agency IplanRIO released Rio-3.5-Open-397B earlier this week, presenting it as an original 397-billion-parameter large language model trained in-house for the city government. The launch drew immediate attention given the enormous compute resources such a model would require, with the city claiming it represented a major Brazilian achievement in AI.
Why it matters
Within days, the team at Nex-AGI published a detailed forensic analysis showing the model is not original at all. Using two completely independent methods, they demonstrated that Rio-3.5 is simply a weighted merge of their own Nex-N2 model and Alibaba's Qwen3.5-397B, blended at roughly 60-40. When the 'You are Rio' system prompt was removed, the model identified itself as Nex 79% of the time and even recited Nex-AGI's internal backstory verbatim. Weight-tensor analysis across all 60 layers confirmed the exact same blend ratio to thousands of standard deviations.
What's next
The incident highlights growing concerns about transparency in AI development as governments and institutions race to claim AI capabilities. Nex-AGI noted that model merging and fine-tuning are legitimate techniques, but presenting a merged model as a ground-up original crosses an ethical line. The Rio-3.5 repository remains on HuggingFace as of publication, and IplanRIO has not yet responded to the findings.