Rio de Janeiro's claimed homegrown LLM appears to be a straightforward blend of two existing models with no original training.

What happened: A 397B-parameter model called Rio-3.5-Open-397B, presented by IplanRIO as an original creation, is claimed by Nex-AGI to be a direct element-wise merge of Nex and Qwen3.5-397B-A17B (approximately 0.6 Nex / 0.4 Qwen) with no evidence of independent training. When Rio's system prompt is removed, the model identifies itself as Nex 79% of the time and recites Nex-AGI's backstory word-for-word. Every weight tensor across all 60 layers matches the 0.6/0.4 blend to thousands of standard deviations.

Why it matters: Claims of homegrown AI development carry weight in public discourse and funding decisions. If a model is presented as original work when it is actually a merge of existing models, it misrepresents both the technical achievement and the resources invested. This touches on transparency in how AI organizations claim credit for their systems.

What to watch: The evidence is presented in two independent ways — behavioral inference from the model's own responses and direct weight-tensor analysis across the entire network architecture. The specificity of the comparison (0.6/0.4 ratio, 60 layers, thousands of standard deviations of match) suggests the claim can be independently verified or contested.