Rio's "homegrown" 397B LLM is a 0.6/0.4 merge — and the HuggingFace page quietly admitted it

A municipal IT department shipped a frontier-sized model

IplanRIO, the City of Rio de Janeiro's IT company, released Rio-3.5-Open-397B, presented as a homegrown Qwen3.5 fine-tune that beat comparable open models on benchmarks. The GitHub issue that became the story argues it's nothing of the kind: a straight weighted merge of ~60% Nex-N2 Pro and ~40% Qwen3.5-397B-A17B, with Nex-N2 having shipped about a week earlier.

The evidence is unusually clean for an attribution fight. Strip Rio's system prompt and the model identifies as "Nex, from Nex-AGI" 79% of the time and "Rio" 0%, reciting Nex-AGI's backstory verbatim. And the weights settle it:

"Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen — across all 60 layers and every component of the network. Other finetunes cannot be explained as interpolations." — from the issue filed by 00INDEX

The model page edited itself mid-controversy

The thread's decisive catch wasn't new analysis. It was a diff. A commenter noticed the HuggingFace page had grown a paragraph that wasn't there before.

"The model is built via a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, proceeded by On-Policy Distillation from a stronger model. We detected an incorrect upload in the previous version, where the base merged version was upload instead of the final distilled model. We are sorry for the confusion and apologize profusely." — a commenter quoting the newly-added paragraph

That reframes the debunk: the merge is now conceded; the dispute is whether the public model was the intended distilled version or, as the page now claims, an accidental upload of the raw blend.

"Can we just merge all the open weight models?"

The more interesting reaction was technical surprise that the trick worked at all. A linear average of two networks usually breaks them; here it apparently helped.

"I find it amazing how robust the current deep learning models are. A simple linear combination of every weight did not degrade the performance of the model, but enhanced it." — hintymad

A Brazilian engineer with no affiliation offered the most generous reconstruction: they likely didn't disclose Nex because Nex is itself a Qwen derivative, the real gain was supposed to come from merge-plus-distillation, and the model went viral organically over Brazil's World Cup weekend before the mayor seized the moment. Another commenter shrugged the whole thing off: Nex is based on Qwen too, so "don't think they're too far off."

What the diff actually proves

This isn't a story about Brazilian fraud, despite the jokes in the thread. It's a story about how little a model can hide now: an independent analyst reconstructed the exact mixing ratio from the weights alone, and the only move left was to edit the README. The merge itself may even be defensible — open weights, Apache-style lineage, a real distillation step that genuinely improved benchmarks. What's not defensible is shipping it as "homegrown" until someone ran the numbers. The lesson for every org tempted to launch a national-pride model: the weights are a confession waiting to be read.