The Rio de Janeiro LLM Merge Scandal and the Theater of AI Sovereignty
A major technical controversy has erupted around the municipality of Rio de Janeiro’s "homegrown" large language model, Rio-3.5-Open-397B. Advertised by the city's IT company (IplanRIO) as a major breakthrough in local AI development with superior Portuguese language capabilities, independent researchers have proved that the model is actually a simple, lazy mathematical weight merge of Nex-N2 Pro and Qwen3.5-397B-A17B.
This incident exposes the growing theater of "AI sovereignty," where municipal and national public funds are being used to rebrand open-weight models as homegrown innovations to capture political prestige and public sector budgets.
The Technical Proof
The creators of the Nex-N2 model (Nex-AGI) provided two independent, mathematically irrefutable proofs that Rio-3.5-Open-397B did not undergo any independent post-training or fine-tuning of its own:
- System Prompt Removal and the "Watermark" Identity: Rio-3.5-Open-397B was shipped with a hard-coded system prompt forcing it to state "You are Rio." However, when this prompt was removed, the underlying model identified itself as "Nex, from Nex-AGI" in 79.2% of test cases and as "Rio" in 0% of cases. It even recited Nex's private, bespoke backstory word-for-word, including references to the "Shanghai Innovation Institute" that had been baked into Nex's training data.
- Weight Tensor Collinearity: A weight merge exhibits a strict linear relationship: $\text{Rio} = \alpha \cdot \text{Nex} + (1 - \alpha) \cdot \text{Qwen}$. Across all 60 layers and every component of the network, the mixing weight ($\alpha$) was recovered as a highly stable $0.571 \pm 0.0016$. More importantly, the collinearity ($\text{cos_fit}$) of the weight tensors was measured at $0.98 - 0.99$. In a multi-billion-dimensional space, two unrelated models would have a collinearity of approximately $0$. A score of $0.99$ across all layers is a statistical impossibility unless Rio's weights were directly blended from Nex's.
The "Incorrect Upload" Defense
Faced with undeniable technical evidence, the Rio de Janeiro team updated their model card on Hugging Face to acknowledge the merge, but attempted to deflect criticism by claiming an "incorrect upload." They asserted that the base merged version was uploaded by accident instead of their "final distilled model."
Commenters on Hacker News noted that this is the standard playbook when public-sector or VC-backed AI projects get caught rebranding open-weight models (recalling the infamous "Reflection 70B" drama).
The Political Economy of "Local AI"
The scandal highlights a growing trend of municipal and regional governments attempting to build a local "AI brand" using public money. While model merging is a highly effective and valid technique in the open-source community to combine the strengths of different fine-tunes, portraying a simple blend as a "homegrown, trained-from-scratch" model represents a form of high-tech fraud. It raises serious questions about how public IT infrastructure funds are monitored and audited in the age of open-weights.