Rio's 397B 'homegrown' open model looks like a Nex-Qwen merge, researchers claim

A GitHub issue filed against IplanRIO's Rio-3.5-Open-397B accuses the city of Rio de Janeiro's headline open model of being a weight-space merge of two existing models rather than an independent training run. The complainant is Nex-AGI, whose own Nex model the issue says was used as one of the inputs alongside Alibaba's Qwen base model.

The team's evidence comes in two flavors. The behavioral test is simple: strip the Rio system prompt and the model identifies itself as "Nex, from Nex-AGI" 79% of the time and as Rio 0% of the time, even reciting Nex-AGI's original backstory verbatim. The numerical claim is stronger: across all 60 layers and every weight tensor, the Rio weights are, by the authors' measurement, "to thousands of standard deviations" the same 0.6/0.4 blend of Nex and Qwen. Normal post-training fine-tunes, they argue, don't look like clean interpolations of two parents in tensor space.

IplanRIO presented Rio-3.5-Open-397B as a sovereign AI effort, which is the part that makes the dispute matter beyond a credit fight. If the analysis holds, this is not a fine-tune attribution problem but a public-money story about a city procuring an "independent" frontier model that wasn't trained.

Why it matters

If you ship products on open weights, this is the second high-profile provenance claim in months and it argues for treating model identity probes and weight-space fingerprints as part of due diligence, not nice-to-have. For anyone procuring sovereign or national AI: ask what was trained, on what compute, and check for tensor-level overlap with available base models before signing the contract.