Rio Grande Model Case Follow-up: Officials Claim Model Loss Led to Forced Retraining, Accused of Embezzling Public Funds Through AI Project

According to Pulse Beat monitoring, the IT company IplanRIO, owned by the Rio de Janeiro municipal government in Brazil, recently issued an official statement in response to the controversy surrounding its earlier model, which was accused of plagiarizing the Chinese open-source project Nex-N2 Pro and Alibaba's Qwen 3.5. In the official apology, the company provided a bizarre excuse, stating that the final model's weight file used for deployment had been lost, forcing them to reschedule the training. The official response failed to quell the public opinion and instead sparked strong suspicions that the project was using the concept of AI to embezzle public funds.

In its statement, IplanRIO admitted that the Rio-3.5-Open-397B was not a self-developed and independently trained base model as advertised but a product of fine-tuning and merging based on existing open-source models. The company also apologized for not previously acknowledging the Nex-AGI team in the documentation. Regarding why the uploaded actual weights turned out to be a "patchwork" linear combination of Nex and Qwen, the company explained that due to an "operational mistake," they uploaded an intermediate baseline version used for comparison, rather than the distilled and fine-tuned final version. However, the company subsequently stated that despite attempts to recover the final model, they were unsuccessful and would need to retrain and externally validate before reposting. Currently, the Rio-3.5-Open-397B page on Hugging Face has been completely taken down.

Previously, the municipal government allocated 500,000 Brazilian reais (approximately $100,000) for the large-scale model project's training budget, but what was ultimately delivered was a zero-training patchwork weight, with no evidence even of the so-called "final version weight." The essence of the incident is an extremely crude embezzlement scam: taking public funds but only using open-source weights for a shoddy deliverable, being caught red-handed, then attempting to cover it up with a claim of "file loss," and finally trying to deceive again with promises of "retraining."