Rio's municipal tech agency, IplanRIO, presented a 397 billion parameter AI as original work, but Nex AGI, the open source AI lab whose model was one of the two, says the trained parameters (weights) are a roughly 60/40 blend of two existing open models — and IplanRIO has since…
Rio de Janeiro's city government says it built a frontier-class AI from scratch. The model's own weights tell a different story: a roughly 60/40 blend of two existing open-source systems, with no evidence of original training behind it.
The model is Rio-3.5-Open-397B, released on Hugging Face by IplanRIO, the information technology arm of the Prefeitura do Rio de Janeiro. The model card describes a 397-billion-parameter mixture-of-experts system with 17 billion parameters active per token, a roughly one-million-token context window, and an MIT license. The card presents the model as a "frontier-class general-purpose AI model developed by IplanRIO," states it was "post-trained from Qwen 3.5 397B," and credits a training-free inference framework called "SwiReasoning." It does not mention Nex-AGI, Nex, or any other upstream model.
That is the public claim. On Sunday, the open-source lab Nex-AGI opened issue #4 on its Nex-N2 GitHub repository, alleging that Rio's weights are not an original build. Nex-AGI's argument rests on two reproducible observations, both of which the lab says any third party with the same base models and a GPU can repeat.
The first observation is behavioral. Rio-3.5-Open-397B ships with a hard-coded system prompt that instructs the model to identify as "Rio." When Nex-AGI stripped that prompt and asked the deployed model who it was, the model answered "Nex, from Nex-AGI" 79 percent of the time and "Rio" 0 percent of the time, and it recited Nex-AGI's own backstory language verbatim. The second is structural. A weight-tensor analysis reported in the issue finds that every parameter in Rio-3.5-Open-397B matches a 0.6 share of Nex-AGI's Nex-N2_pro and a 0.4 share of Qwen3.5-397B-A17B to thousands of standard deviations, across all 60 layers and every network component. The issue states that other public fine-tunes of the same Qwen base cannot be explained as interpolations of these two models in the same way.
A weight merge, or weight-space merge, is a recognized technique: the parameters of two existing models are combined, element by element, to produce a single new set of weights. Many open-source projects publish merges openly, and the technique itself is not the problem here. The problem is that a city government presented a 60/40 blend as a homegrown 397-billion-parameter model without disclosing the upstream lineage. The constructive frame is provenance, not theft: the gap between the "homegrown" framing and the technical reality on disk is the public-interest story, and the merge is the mechanism, not the scandal.
Nex-AGI is not a neutral party. Its Hugging Face organization page describes it as "an innovation alliance initiated by the Shanghai Innovation Institute," and it has a commercial interest in how its Nex-N2 weights are attributed. The technical evidence is independently reproducible, and Nex-AGI's stake is worth flagging once. The claim does not stand or fall on the accuser's motives: any team with Nex-N2_pro and Qwen3.5-397B-A17B can run the same arithmetic against the Rio weights on disk and reach their own conclusion.
Update — June 14, 2026: The Rio-3.5-Open-397B model card has been updated since Nex-AGI's issue was published. The revised card now states: "The model is built via a merge of nex-agi/Nex-N2-Pro and Qwen/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 uploaded instead of the final distilled model. We are sorry for the confusion and apologize profusely."
This acknowledgment resolves the core factual dispute Nex-AGI raised: the merge is now confirmed by both parties. The remaining questions are narrower — whether the "final distilled model" differs materially from the publicly posted weights, and whether the original model card's "developed by IplanRIO" framing adequately represented what was actually released. IplanRIO and Prefeitura do Rio did not respond to a request for comment on this story before publication. The GitHub issue is less than a day old, and the factual picture may shift if the city agency publishes further detail, if a third party reproduces or refutes the tensor analysis, or if the model card is amended.
For cities, federal agencies, and any public body publishing "we built an AI" claims, the Rio case is a clean test of a missing norm. Provenance disclosure — what a model is built on, what was added in fine-tuning, and what, if anything, was trained from scratch — has no standard format and no enforcement. Procurement officers, RFP reviewers, and journalists covering municipal AI will keep running into the same gap until the disclosure itself becomes part of what "developed by" means. Rio-3.5-Open-397B is the first reproducible public example of how that gap reads at scale.