The city's municipal IT company fine-tuned an open-source Alibaba model to beat DeepSeek V4, but a weight-merge controversy is already clouding the achievement.
A city government just out-benchmarked one of the most talked-about AI labs on the planet. Rio de Janeiro’s municipal IT company, IplanRIO, released a 397 billion parameter AI model that scored higher than DeepSeek V4 Pro on a key benchmark, and it did so by building on top of Alibaba’s open-source foundation rather than training anything from scratch.
The model, called Rio 3.5 Open 397B, went live on Hugging Face on June 13 under an MIT license. It scored 70.8 on Terminal-Bench 2.1, compared to 67.9 for DeepSeek V4 Pro and 52.5 for the base Qwen model it was derived from.
How a city government built a top-tier AI model #
Rio 3.5 is a fine-tuned version of Alibaba’s Qwen 3.5-397B-A17B, which uses a Mixture-of-Experts architecture. The model has 397 billion total parameters, but only 17 billion are active at any given time.
The model also integrates what IplanRIO calls the SwiReasoning inference framework, and supports a 1 million token context window.
This release follows the Rio 3 Open model family, which IplanRIO introduced in April 2026. The broader strategy is explicitly about what the city calls “technological sovereignty,” the idea that municipalities shouldn’t be entirely dependent on a handful of Silicon Valley companies for their AI infrastructure.
Rio has reportedly committed $550 million to data center investments as part of its larger “AI City” initiative.
The controversy nobody’s ignoring #
Not everyone is buying the narrative of homegrown innovation. Almost immediately after the release, an AI outfit called Nex-AGI claimed that Rio 3.5’s weights were essentially a straightforward 60/40 blend of Nex-AGI’s own model and the base Qwen model, with no independent training involved.
IplanRIO responded by labeling the claims an “operational error.” That phrasing is notably vague, neither a full denial nor an admission. It leaves open the question of whether some weight merging occurred but was unauthorized, or whether Nex-AGI’s claims are simply wrong.
If independent researchers confirm the weight-merge allegations, it would significantly undermine Rio’s credibility in the AI community and cast doubt on whether the benchmark results represent genuine capability improvements. The difference between “we fine-tuned a model with novel techniques” and “we blended two existing models together” is the difference between innovation and arbitrage. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our