← Back to feed

Rio de Janeiro’s city government announced Rio3.5 as a homegrown large language model, releasing benchmark figures that claimed it outperformed Alibaba’s Qwen3.7-Max. The announcement positioned the model as a milestone for Brazilian sovereign AI — a city-level success story in a field dominated by US and Chinese frontier labs.

The model appears to be a merge of existing open-source weights.

Analysis published on GitHub exposes the architecture: Rio3.5 is assembled from pre-trained open-source components using model merging techniques, which combine existing weight matrices without any novel training data or compute investment. The city conducted no independent pretraining. It produced no model card documenting a training run, dataset, or compute budget. The benchmarks cited were self-reported.

What Model Merging Is — and Is Not

Model merges are a legitimate and widely practised engineering technique. Developers take two or more pre-trained open-source models, combine their weights using methods like SLERP or TIES merging, and sometimes get a result that scores better on certain benchmarks than either source model alone. The technique requires no GPU cluster, no novel data, and no research contribution. It requires a laptop and a weekend.

The result is not a trained model. It is a weighted interpolation of someone else’s trained models, typically licensed under Apache 2.0 or similar open-source terms that do permit downstream merges. The problem is not the technique. The problem is describing the output as a homegrown model.

Rio3.5’s claimed benchmark edge over Qwen3.7-Max is plausible given how merges work — a narrow combination of Qwen weights plus another open-source base can score higher than either source on specific tasks, particularly if the benchmark categories are selected to favour the merge. This is benchmark optimisation, not AI capability development.

A Pattern Across Government AI Announcements

Rio3.5 is not an isolated case. Over the past eighteen months, multiple national and municipal governments have announced “homegrown” AI models that, on inspection, are fine-tunes or merges of openly licensed weights from Meta’s Llama family, Alibaba’s Qwen series, or Mistral. The underlying models are often publicly disclosed in the system card or technical documentation — when such documents exist at all.

The incentive structure is clear. Announcing sovereign AI development generates political capital, attracts public funding, and positions a government as an AI-capable actor in an increasingly strategic field. The actual work of building a frontier model from scratch — requiring hundreds of millions of dollars, months of compute time, and dozens of ML researchers — does not produce a press release on a city government’s timeline.

What results is an arms race in government AI theater: announcements that are technically accurate (a new model has been deployed) but materially misleading (the development effort was minimal).

Why Benchmark Numbers Travel Without Context

Rio3.5’s claim against Qwen3.7-Max will circulate in media coverage that does not distinguish between self-reported benchmarks on cherry-picked tasks and independent evaluation on standardised leaderboards. A model merge that scores well on one Portuguese-language legal document task can be reported as “beating Qwen3.7-Max” without the qualifier that the comparison is local, self-referential, and not replicated on SWE-bench, LiveBench, or any arena with independent judges.

The disclosure came from GitHub — a researcher examining the model’s weights and filing an issue, not a government audit or a third-party evaluation body. There is no institutional mechanism for verifying government AI capability claims before they enter public circulation.

Rio3.5’s benchmark numbers have not been reproduced by an independent evaluator. The model’s architecture and training lineage have not been publicly documented by the city. The source models it was merged from have not been disclosed.

The benchmark claim is disputed. The model origin is not.