Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model

	Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model(github.com)
	366 points by unrvl22 21 hours ago \| 193 comments
	tl;dr: Rio de Janeiro's IplanRIO claims its Rio-3.5-Open-397B is an original homegrown LLM, but Nex-AGI provides evidence it's actually a static element-wise weight merge (~60% Nex, 40% Qwen3.5-397B-A17B base) with no apparent additional training. When the hardcoded system prompt is stripped, the model identifies itself as "Nex" 79% of the time and reproduces Nex-AGI's backstory verbatim, while tensor-level analysis confirms a uniform 0.6/0.4 blend across all 60 layers.
	HN Discussion: ~Offers an alternative explanation that the merge was intentional and undisclosed distillation was the missing piece ↑Marvels at the technical robustness of deep learning models surviving linear weight blending ↑Confirms and summarizes the article's findings about the undisclosed merge ↓Dismisses the issue, suggesting similar practices are rewarded elsewhere without scrutiny ↑Agrees this undermines trust and calls for better provenance and transparency standards