The Weights Do Not Lie: How Rio's 'Sovereign' AI Was Unmasked by a Laptop Audit
The Laptops of Copacabana
When the municipal IT department of Rio de Janeiro, IplanRIO, announced the release of Rio-3.5-Open-397B on Hugging Face, the accompanying press release read like an epochal shift in Latin American technology. Here, supposedly, was a “frontier-class municipal AI model” developed in-house by local public servants, fine-tuned from Alibaba’s Qwen 3.5 base to beat state-of-the-art open models on global benchmarks. Political leaders and municipal pride celebrated a sovereign triumph. There was only one problem: the model kept insisting, under direct questioning, that it was actually “Nex,” an agent framework trained thousands of miles away by a commercial startup called Nex-AGI.
The model had literally outed itself. In over 70% of identity trials stripping away municipal system prompts, the checkpoint answered that it was built by the “Shanghai Innovation Institute.” It turns out that Nex’s developers had watermarked their model with comprehensive self-identity prompts during training. When Rio’s municipal IT department downloaded Nex’s weights and ran a basic linear merge, those deep watermarks remained intact. It is the AI equivalent of a high-school student duplicating a classmate’s essay, changing the name on the cover page, but forgetting to delete the original student’s name hidden in the footer.
The Incontestable Math of the Range Request
The technological heist was proven not with speculative gossip, but through elegant linear algebra. Because the 397-billion-parameter Mixture-of-Experts (MoE) file is a massive, multi-hundred-gigabyte download, ML researcher Tiezhen Wang bypassed downloading the full weights. Instead, using Python’s NumPy and HTTP range requests, he extracted only the byte offsets for specific multi-million-dimensional MoE gating vectors and layer norms directly from the Hugging Face repository.
The logic of a weight merge is mathematically deterministic. If Rio is a simple linear blend of Nex and Qwen, then:
$$\mathbf{W}{\text{Rio}} = \alpha \mathbf{W}{\text{Nex}} + (1 - \alpha) \mathbf{W}_{\text{Qwen}}$$
This means the deviation vector of Rio’s weights from the Qwen base must point in the exact same direction as Nex’s deviation from Qwen:
$$\mathbf{W}{\text{Rio}} - \mathbf{W}{\text{Qwen}} = \alpha (\mathbf{W}{\text{Nex}} - \mathbf{W}{\text{Qwen}})$$
By analyzing the cosine similarity between these difference vectors ($\cos_fit$), Wang proved that the direction matched at an astronomical $0.99$ confidence level. In a 2-million-dimensional space, the probability of two independently trained models sharing a deviation vector direction is virtually zero. The scaling factor $\alpha$ consistently clustered at $\approx 0.57$. The municipal “frontier model” was, with mathematical certainty, nothing but a $57%$ Nex-N2-Pro and $43%$ Qwen3.5 weight merge executed via a free command-line tool.
| Model Component Analyzed | Interpolation Factor ($\alpha$) | Cosine Fit Similarity ($cos_fit$) |
|---|---|---|
| MoE Router (Layer 0) | 0.573 | 0.992 |
| MoE Router (Layer 15) | 0.647 | 0.962 |
| MoE Router (Layer 30) | 0.627 | 0.967 |
| MoE Router (Layer 59) | 0.567 | 0.997 |
| Shared Expert Gates (Layer 0) | 0.568 | 0.997 |
| Shared Expert Gates (Layer 30) | 0.581 | 0.988 |
Domestic Pride vs. Algorithmic Plagiarism
When confronted with the GitHub issue detailing the linear algebra, the municipal team’s reaction was a classic study in prompt damage control. They quietly updated their Hugging Face model card to include a frantic correction: “We detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model.” They edited the repository to acknowledge the merge but claimed it was merely a preliminary base for “On-Policy Distillation.”
Among developers and observers, the reaction split along lines of geopolitical sympathy and deep-tech cynicism. On Hacker News, some locals jumped to defend the municipal workers. Reader thimabi commented: “I wouldn’t describe what happened here as incompetence. As a ‘carioca’ [Rio resident], I am pleasantly surprised to know that the government’s IT department is involved in AI work — even without the budget to create its own models from scratch.” Meanwhile, others questioned why a municipal government was doing this at all, with mrandish asking: “Because… lack of a good open-weight LLM is a pressing need high on the municipal priorities list for Rio de Janeiro citizens?”
The Era of Sovereign Model Washing
The Rio de Janeiro saga exposes a grander trend: the rise of “Sovereign AI Washing.” Because tools like mergekit have democratized weight interpolation to a point where any hobbyist can blend two models on a high-end consumer laptop, we have entered an era of deep bureaucratic vanity. Municipalities, federal departments, and defense contractors—eager to show they are “leading the AI revolution” without the billions of dollars required for native compute and pre-training—can download a model from Alibaba or Meta, blend it with a startup fine-tune, and wrap it in the national flag.
In this new ecosystem, benchmarks are easily gamed and press releases practically write themselves. But as the mathematical audit of Rio-3.5 proves, the weights do not lie. No matter how many layers of public relations, municipal registries, and local licensing are draped over the weights, the linear alignment of the vectors remains a digital fingerprint. If your sovereign technological future is built entirely on another builder’s foundation, the math will eventually expose the scaffolding.