Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
Points and comments are a snapshot, not live.
Rio de Janeiro's Rio-3.5-Open-397B model is a 60/40 weighted merge of Nex-N2 Pro and Qwen3.5-397B, not an original training effort.
Nex-AGI researchers demonstrate that Rio-3.5-Open-397B, presented by IplanRIO as a homegrown achievement, is a direct element-wise merge of their Nex-N2 Pro model (60%) with Qwen3.5-397B-A17B (40%). Two independent lines of evidence support this: first, when the model's hard-coded "You are Rio" system prompt is removed, it identifies itself as "Nex, from Nex-AGI" 79% of the time and as "Rio" 0% of the time, while reciting Nex's proprietary backstory word-for-word. Second, every weight tensor across all 60 layers shows the same 0.6/0.4 blend ratio with collinearity measurements near 1.0, indicating a fixed mathematical merge rather than independent post-training. The researchers conclude Rio contains no evidence of any original training.
What commenters are saying
The thread splits between those focusing on attribution and fraud (claiming misuse of public funds, false capability claims) and those contextualizing model merging as a known technique in open-weight communities. Several commenters note that Rio's HuggingFace page was updated to acknowledge the merge and claim an accidental upload of the wrong file, though no corrected version has appeared. One comment questions whether GitHub Issues are the right venue for this type of accusation. Others observe that Nex-N2 itself derives from Qwen, making the merge technically feasible, and compare this to common practices in Stable Diffusion communities.