Rio City’s Domestic LLM Suspicion Exposed as Merge of Existing Models, Whistleblower Reveals with Hard Evidence

The origin of the large language model “Rio-3.5-Open-397B,” recently released by Rio de Janeiro’s public IT company IplanRIO, is now under serious suspicion. On June 14, 2026, AI startup Nex-AGI posted on its own GitHub Issue that Rio-3.5 is merely an element-wise combination of the weights of Nex-AGI’s model “Nex-N2-Pro” and “Qwen3.5-397B-A17B” developed by the Qwen team under Alibaba Group.

Background of the Accusation

IplanRIO had announced Rio-3.5-Open-397B as a “397B parameter model independently trained from scratch.” It attracted attention as a domestic LLM project involving the city’s technology bureau, but Nex-AGI presented evidence that flatly denies this claim.

According to Nex-AGI, the weight matrices of Rio-3.5 across all 60 layers and all network components can be explained as a linear combination of approximately 0.6 times Nex-N2-Pro and 0.4 times Qwen3.5-397B-A17B. This ratio matches with a precision of thousands of standard deviations, which cannot be attributed to coincidence or ordinary fine-tuning.

A Model That Calls Itself “Nex”

The most striking part of the accusation is the model’s response when asked about its own identity. Nex-AGI removed the hardcoded system prompt “You are Rio” from Rio-3.5 and asked the model to introduce itself. As a result, 79% of the time it answered “I am Nex of Nex-AGI,” while it never answered “I am Rio.”

The accuser also noted that “Rio-3.5 recites the unique backstory of Nex-AGI’s organization verbatim.” While fine-tuning can change a model’s behavior, it “cannot explain” the internal weight matrices being a simple weighted average of existing models.

The accusation was published as Issue #4 on Nex-AGI’s GitHub repository “Nex-N2.” Within hours of publication, it received 226 points and over 123 comments, and became a top story on Hacker News.

Technical Verification Details

Nex-AGI presented two independent analysis methods as evidence.

First, behavioral analysis of model outputs. The aforementioned system prompt removal test showed the model’s self-identification is biased toward Nex. This is considered proof that the model has not learned a unique identity as Rio.

Second, statistical analysis of weight tensors. It showed that all weight vectors of Rio-3.5 can be explained as a linear interpolation of Nex-N2-Pro and Qwen3.5-397B. Deviations from the specific ratio are within measurement error, and there is no trace of additional training.

Nex-AGI claims that “fine-tuning other models does not result in such a simple linear interpolation.” Fine-tuning involves gradient updates to specialize for specific tasks, so weight changes typically show different patterns per layer or channel. The uniform ratio mixing across all parameters strongly suggests an intentional merging process, not original training.

The Credibility Issue of Open Source LLMs

This incident once again highlights the problem of model origin and credibility in the open source LLM community. Developing large language models requires enormous computational resources and datasets, and it is impossible to take at face value claims from organizations that say they trained a model from scratch.

IplanRIO had positioned this project as a “domestic model led by a Brazilian public institution.” The lack of transparency in public procurement and government AI projects concerns not only technical validity but also taxpayer fund usage and governance.

A similar case occurred in 2023 when Stanford University released the “Alpaca” model, which was based on Meta’s LLaMA. However, Alpaca was clearly disclosed as “fine-tuned from LLaMA,” unlike this case where it was falsely presented as training from scratch.

IplanRIO’s Response

As of writing, no official comment has been released by IplanRIO or the city of Rio de Janeiro. The Rio-3.5-Open-397B repository on GitHub remains public, but numerous questions have been raised from the community following Nex-AGI’s accusation.

All eyes are on whether IplanRIO will provide an explanation or retract the model. If Nex-AGI’s claims are correct, the city’s announcement of proprietary training contains false information, potentially affecting not only technical credibility but also the trustworthiness of the administration.

Technical Background of Model Weight Merging

Weight merging of large language models itself is a research topic for model compression and knowledge integration from multiple models. Representative methods include Model Soups, TIES-Merging, and DARE, which mix weights of multiple fine-tuned models at specific ratios to improve task performance.

However, these methods are inherently techniques premised on additional training, not intended for faking a model trained from scratch. This case shows that a “proprietary model” can be generated through simple weighted averaging, forcing the entire industry to reconsider verification processes for model releases.

Editorial Opinion

This accusation is a crucial case for maintaining the health of the open source LLM ecosystem. The two independent pieces of evidence presented by Nex-AGI are not mere suspicion but statistically solid. If IplanRIO fails to provide a sufficient explanation, the credibility of AI projects led by public institutions will suffer a serious blow.

In the short term, this case may raise the bar for LLM releases by other government and public IT organizations. By visualizing the risk of calling a non-originally-trained model “domestic,” projects with low transparency will face stricter community scrutiny.

From a long-term perspective, technical methods for verifying model origin—such as statistical analysis of weight tensors and system prompt removal tests—may become established as standard auditing processes. As open source LLMs take on roles as social infrastructure, mechanisms to verify their origin and training data transparency become essential.

We call on IplanRIO to provide the model’s weights and training logs for verification by a third party. To restore trust in the technical community, not just explanations but concrete evidence disclosure is needed. This case also strongly suggests the importance of technical due diligence in AI public procurement, and it is urgent to establish standards for when administrative bodies introduce or release external AI models.

References

• GitHub Issue: Rio-3.5-Open-397B ≈ 0.6 x Nex-N2_pro + 0.4 x Qwen — Published June 14, 2026
• Hacker News Discussion Thread — Published June 14, 2026
• Cloudflare acquires VoidZero to boost Vite/Astro development — Posted on this site
• AMD Helios MI455X faces performance challenges with Ethernet connection — Posted on this site