79% of the time, with its hard-coded "You are Rio" system prompt removed, the model identifies itself as "Nex, from Nex-AGI". It identifies as "Rio" 0% of the time. That alone would be damning, but nex-agi also published the arithmetic: every weight tensor in Rio-3.5-Open-397B is, to thousands of standard deviations, a 0.6/0.4 element-wise blend of their Nex-N2_pro and the official Qwen3.5-397B-A17B base.
The 0.6/0.4 Interpolation
Rio-3.5-Open-397B was presented by IplanRIO as an original 397B model trained from scratch or heavily fine-tuned. The GitHub issue from nex-agi shows the math is simpler: take Nex-N2_pro, mix in 0.4 of Qwen, and you get Rio's weights. This holds across all 60 layers and every network component. Other finetunes can't be explained as interpolations, the authors note. The claim is not just plausible - it's mathematically overdetermined.
The "You Are Rio" System Prompt Test
Nex-AGI's second proof is behavioral. With the custom system prompt stripped, Rio's own deployed model repeatedly outputs Nex-AGI's bespoke backstory word-for-word. It never claims to be Rio. That's a fingerprint no amount of weight averaging can hide. For an organization that branded this as a "homegrown" LLM, the model's own behavior disagrees.
What This Means for Open Model Claims
Model merging is a legitimate technique - but presenting a merged model as an original training effort is not. The evidence here is specific and reproducible: exact blend ratios, layer-by-layer tensor comparison, and behavioral self-identification. Expect more scrutiny on how municipal AI projects disclose their model origins. If you publish a model, the weights will tell the truth.
Whether IplanRIO responds or not, the technical record is now clear. Next time someone claims a novel 397B model, check the tensors first.
Source: Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
Domain: github.com
Comments load interactively on the live page.