Follow up to the Rio Big Model Shell Scandal: Officials argue that the model was lost and forced to retrain, accused of embezzling public funds through AI projects According to Beating monitoring, IplanRIO, an IT company under the Rio de Janeiro Municipal Government in Brazil, recently issued an official statement in response to the controversy over the previous model being accused of being a shell of the Chinese open-source project Nex-N2 Pro and Alibaba Qwen 3.5. While apologizing, the official also provided a bizarre defense, stating that the final model weight file used for publication has been lost and forced to reschedule training. The official response failed to quell public opinion and instead sparked strong doubts about the project's alleged use of AI concepts to embezzle public funds. IplanRIO acknowledges in its statement that Rio-3.5-Open-397B is not a self-developed and trained base model as advertised, but rather a product of fine-tuning and merging based on existing open source models, and apologizes for not thanking the Nex AGI team in the document earlier. The official explanation for why the actual weights uploaded were the "stitched version" of Nex and Qwen's linear merge is that due to an "operational error", an intermediate baseline version was uploaded for comparison, rather than the final version after distillation and post training. But the official immediately stated that the team had attempted to recover the final model but was unsuccessful, and it must be retrained and externally validated before it can be re released. At present, all pages of Rio-3.5-Open-397B on Hugging Face have been taken down and cleared. The Rio Municipal Government had previously allocated 500000 Brazilian reals (approximately 100000 US dollars) in training funds for the large model project, but in the end, a zero training stitched weight was handed over, and even the so-called "final version weight" was found to be unfounded. The essence of the incident is an extremely crude scam to embezzle funds: taking public funds but only using open source weights to stitch and submit them, being caught on the spot and covered up with "lost files", and finally using "retraining" to deceive. [Original link]