The city of Rio de Janeiro’s claim to have built a 397-billion-parameter AI model has unraveled in spectacular fashion. Open-source developers quickly discovered that the so-called self-developed model, Rio-3.5-Open-397B, is nothing more than a direct linear merge of two existing Chinese open-source models — with zero additional training.
The model’s weight files were found to be a simple 60/40 blend of Nex-AGI’s Nex N2 Pro and Alibaba’s Qwen 3.5, stitched together without any post-training or fine-tuning. It's essentially the AI equivalent of claiming you built a car from scratch when you just bolted together two off-the-shelf engines.
The exposed model was supposed to showcase Brazilian tech independence, but instead it highlights the perils of trying to cut corners in the open-source world, where every line of code and weight is scrutinized by a global community of volunteers.