Decrypt via Reddit

IplanRIO Rio 3.5 Exposed as Nex-Qwen Weight Merge

alibaba china ai open source sovereign-ai model-weights

Key insights

  • Nex-AGI proved Rio 3.5 is approximately 60% Nex N2 Pro and 40% Qwen 3.5, with collinearity of 0.993 across all 60 model layers.
  • The model identified itself as 'Nex' 79.2% of the time with its system prompt removed, and zero percent as 'Rio'.
  • IplanRIO blamed the mislabeled release on an incorrect upload, claiming the intended final distilled model was never published.

Why this matters

The R$500,000 public investment framing made Rio 3.5 a political claim, not just a technical one, turning attribution into an accountability issue for municipal AI spending. Nex-AGI's forensic methodology, measuring collinearity across all 60 model layers, gives any open-source developer a reproducible tool to challenge sovereign AI claims from public institutions globally. For founders and AI leaders building on open foundations, this case shows open-source license compliance is not the full standard: public-sector actors face community and reputational enforcement that legal licensing alone cannot absorb.

Summary

IplanRIO, Rio de Janeiro's municipal IT agency, released Rio 3.5 Open 397B on June 13, claiming it was publicly funded, independently developed AI that outperformed DeepSeek v4 Pro and Alibaba's Qwen 3.7 Plus. Mayor Eduardo Cavaliere amplified the announcement the following day: "Today, the world is talking about an open AI model trained in Rio." Shanghai-based Nex-AGI proved it was a direct weight merge: approximately 60% Nex N2 Pro and 40% Qwen 3.5. Collinearity measured 0.993 across all 60 model layers, consistent with deliberate blending rather than original development. Stripped of its system prompt, the model called itself "Nex" 79.2% of the time and zero percent "Rio." Essentially: (IplanRIO, Nex-AGI) are at the center of a public attribution dispute over what counts as sovereign AI development. - IplanRIO updated the model card, blaming "an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model." - Both base models carry open licenses, making the merge legal, but Nex-AGI's position is clear: "In the open-source world, attribution matters." - This follows Cursor's Composer 2 being found built on Moonshot's Kimi K2.5 without disclosure earlier in 2026, marking a recurring pattern of undisclosed model origins. Public institutions framing commercial open-source merges as publicly funded sovereign AI raises urgent questions about how governments account for AI investment.

Potential risks and opportunities

Risks

  • IplanRIO's R$500,000 public investment claim faces scrutiny: if the final distilled model does not exist or differs minimally from the merge, it represents a potential misuse of public funds.
  • Other municipal and national governments that have made similar sovereign AI announcements could face retroactive collinearity audits using Nex-AGI's published methodology.
  • Open-source developers may add stricter attribution-enforcement clauses to future model licenses if institutional misrepresentation by public-sector actors continues, fragmenting the permissive licensing ecosystem.

Opportunities

  • Nex-AGI gains significant reputational capital as the lab whose Nex N2 Pro powered a globally covered sovereign AI release, with their 'We are flattered' public response turning the incident into organic brand awareness.
  • AI provenance and model audit services can position collinearity-based weight analysis as a standard compliance tool for governments and institutions deploying open-source models.
  • The Cursor/Kimi K2.5 and IplanRIO/Nex-AGI incidents in 2026 create an opening for open-source consortiums to push for standardized model origin disclosure requirements on major hosting platforms, including Hugging Face.

What we don't know yet

  • Whether IplanRIO's intended 'final distilled model' actually exists and differs materially from the raw 60/40 Nex-Qwen merge.
  • What municipal or national oversight approved the 'publicly funded sovereign AI' framing for a R$500,000 model release.
  • Whether Nex-AGI or Alibaba will pursue any formal action beyond public disclosure, given both base models carry open licenses.

Originally reported by Decrypt

Read the original article →

Original headline: Rio de Janeiro Claims 'Sovereign' 397B AI Model Beat DeepSeek — Then Quietly Admits It Was a Model Merge of Nex-N2-Pro and Alibaba Qwen 3.5