Rio de Janeiro Releases AI Model, Faces Ownership Claim

IplanRIO, Rio de Janeiro's municipal IT company, published Rio-3.5-Open-397B on Hugging Face under an MIT license, presenting it as a government-developed 397-billion-parameter Mixture-of-Experts model. The model card credited IplanRIO with an approach called SwiReasoning, described as switching dynamically between chain-of-thought and latent-space reasoning using entropy-based signals. Benchmark claims showed competitive performance against DeepSeek and Alibaba's Qwen 3.7 Plus, including scores on Terminal-Bench and SWE-Bench Multilingual.

Nex-AGI, the company behind the open-source Nex-N2-Pro model, published an analysis alleging that Rio-3.5-Open-397B is not an original post-training run but a direct weight merge. Per Decrypt and SquaredTech reporting, Nex-AGI found every weight tensor in the model matches a blend of approximately 60% Nex-N2-Pro and 40% Alibaba's Qwen 3.5 across all 60 layers, with no anomalies. Nex-AGI also reported that when the custom system prompt supplied by IplanRIO is removed, the model self-identifies as "Nex, from Nex-AGI" in 79% of responses. Weight merging is a recognized technique for combining trained models via linear interpolation of parameter tensors; it requires no compute-intensive retraining, and attribution obligations depend on the terms of the source models.

The episode illustrates a verifiable detection gap: benchmark-level performance can be staged using merged weights without disclosing source provenance. Independent verification - weight-level comparison against candidate source models, model card audits, and licensing reviews - is the primary due-diligence tool for evaluating third-party or government-claimed model releases. That the release came from a public institution rather than a private vendor adds accountability stakes: government AI claims may influence procurement, policy, and public trust in ways that vendor announcements do not.

No formal response from IplanRIO has been reported. Signals to monitor:

• •whether IplanRIO publishes training logs or weight provenance documentation rebutting the Nex-AGI analysis
• •licensing implications if Nex-N2-Pro's or Qwen's terms govern redistribution and modification of merged weights under the MIT license IplanRIO applied
• •whether the incident accelerates calls for model provenance standards in government AI procurement