A municipal government claimed a frontier model. It wasn't one.
On June 13, 2026, IplanRio — the technology company owned by the city of Rio de Janeiro — uploaded Rio-3.5-Open-397B to Hugging Face under an MIT license, describing it as a state-of-the-art, 397-billion-parameter system built in-house by a municipal government. Within a day, Nex-AGI, the developer of the open-weight Nex-N2 Pro model, filed a GitHub issue laying out weight-level evidence that Rio 3.5 was not trained from scratch at all: every one of its 60 tensor layers matched a linear interpolation of roughly 60% Nex-N2 Pro and 40% Qwen3.5-397B-A17B, with statistical deviation "thousands of standard deviations" below noise — an exact blend, not a coincidence (GitHub issue #4). Strip out Rio's custom system prompt and the model identified itself as "Nex, from Nex-AGI" in 79% of responses, reciting Nex's internal backstory verbatim.
IplanRio did not dispute the analysis. It updated the model's README to acknowledge the architecture is "a merge combining Nex-N2-Pro and Qwen3.5-397B-A17B, followed by on-policy distillation," apologized for what it called a publishing error — uploading the raw merged checkpoint instead of the finished distilled model — and pulled its benchmark claims against DeepSeek and Alibaba's own Qwen line (Hugging Face model card). The agency later said the project cost roughly R$500,000 (about $100,000), a fraction of what an off-the-shelf enterprise AI deployment would run, and that it never claimed to have trained a foundation model from zero (EpicGeek).
The steelman: public money, public trust, public claims
There is a real case for taking this seriously beyond a coding-forum spat. IplanRio is not a startup optimizing for hype; it is an arm of an elected government, and its initial framing — a homegrown "frontier-class" model beating international rivals — was cited within hours by outlets and social accounts as evidence Brazil had produced sovereign AI capability. That claim shaped a public narrative about municipal technical capacity using taxpayer resources, and it briefly stood unchallenged before the community caught it. A government publisher trades on a credibility premium a hobbyist repo doesn't have; overstating originality — even inadvertently — spends public trust in a way that has consequences the next time IplanRio, or another Brazilian state entity, publishes a technical claim. That the correction came from a rival company's engineers rather than any internal or external review process is itself a fair criticism of IplanRio's quality control before publication.
Why the fix is disclosure norms, not a pre-clearance regime
But the technique at the center of this — merging open-weight models and distilling from a stronger teacher — is not a violation of anything; it is one of the most common and legitimate practices in open-source AI development, explicitly enabled by the MIT and Apache-style licenses both Qwen3.5 and Nex-N2 Pro were released under. Community model-merging tools (mergekit and its peers) exist precisely because recombining public weights is a cheap, effective way for smaller teams — including public-sector ones without hyperscaler budgets — to field a capable model. Treating the merge itself as the offense would criminalize exactly the kind of low-cost, high-leverage experimentation that lets an agency spend $100,000 instead of the tens of millions a from-scratch frontier run requires.
The actual failure is attribution: IplanRio's initial model card credited only Qwen and omitted Nex-N2 Pro entirely, and its marketing implied originality the architecture didn't have. That is a disclosure problem, and disclosure problems have a cheaper, faster remedy than statute — which this episode just demonstrated in real time. Nex-AGI's engineers published forensic weight analysis and got a correction within roughly 48 hours, without a regulator, a court, or a compliance filing. That is the open ecosystem's peer-review mechanism working exactly as intended, and it moved faster than any pre-publication government audit board plausibly could.
Brazil's pending AI framework, PL 2338/2023 — approved by the Senate on December 10, 2024 and now before the Chamber of Deputies — already leans this direction: it requires "significant-risk" AI systems to carry accessible documentation of capabilities and limitations rather than mandating ex ante approval before release (Senado Federal, PL 2338/2023). That is the right level of intervention. Extending it, or any future rule, to require government pre-clearance of open-weight releases — on the theory that this incident proves agencies can't be trusted to self-report — would slow exactly the kind of cheap public-sector experimentation Rio was attempting, while adding little the community didn't already deliver for free. What Brazilian regulators should take from Rio 3.5 is narrower: government publishers of AI systems should be held to existing transparency norms — accurate model cards, license-chain disclosure, no benchmark claims that aren't reproducible — with consequences for agencies that get it wrong, not a new licensing gate for the open models that make projects like this possible at all (Canaltech).
Rio 3.5's real lesson is unglamorous: the state should be a careful publisher, not a uniquely regulated one. The community that builds and merges open models proved in under two days it can police provenance claims on its own.