No, I’m not kidding. At first glance, this might sound like one of those AI headlines that has to be fake. But that’s not the case. In fact, the city of Rio de Janeiro in Brazil has released a pretty decent AI model, even outperforming some of the top LLMs in the world.
At a time when the AI race is dominated by Big Tech and well-funded startups, one of the latest high-profile AI models has emerged from a rather unexpected place. Imagine waking up to the news that the Municipal Corporation of Delhi (MCD) has released a new AI model that outperforms some versions of ChatGPT and Claude. That’s more or less what happened in Brazil.
The Rio de Janeiro City Government has released Rio 3.5 Open 397B, a 397 billion parameter AI model. This is suddenly gaining traction across the global AI community. The model was developed by IplanRIO, the city’s IT company responsible for the city’s digital infrastructure and public services, and released on Hugging Face under the MIT license.
Even more impressive than the release is the model’s scorecard in the benchmarks. This outperforms some of the top open source models such as DeepSeek V4, and even some closed models such as ChatGPT and older versions of Claude.
So how did the city council make it happen? Well, Rio 3.5 is based on Alibaba’s Qwen, which was open source until recently.
Rio 3.5 Open 397B is built on Alibaba’s Qwen 3.5-397B-A17B, one of the most capable open source foundation models in China. In other words, IplanRIO did not train a completely new model from scratch. Instead, we took the existing Qwen model, tweaked it, added our own optimizations and an inference framework called SwiReasoning, and released the improved version as an open model.
The day after the US banned the use of Top Claude by foreigners
How Rio approaches AI is also what makes this story so interesting. Especially in light of the recent US ban on foreigners from using top Anthropic Claude AI models such as Mythos and Fable 5. The point of the debate isn’t whether the Rio 3.5 outperforms another model on a benchmark. This is about the fact that the city government was able to take an open source underlying model and turn it into something that is being discussed alongside some of the world’s leading AI systems.
And this may be a glimpse into the future of AI development. Training a frontier AI model from scratch requires significant funding, computing power, and talent. Most governments, businesses, and institutions can’t afford to do that. However, tweaking an open source model that already works is a much more realistic option.
If Rio’s approach is successful, more governments and organizations may follow the same strategy. Rather than relying entirely on U.S. AI companies, they may choose to take a strong open source model and adapt it to local languages, regulations, and use cases.
And that possibility could worry U.S. tech companies. Companies like OpenAI, Anthropic, and Google are increasingly putting their cutting-edge models behind proprietary APIs, while Chinese companies like Alibaba and DeepSeek are aggressively pushing open-weight models. The more capable these models become, the easier it will be for others to build customized AI systems on top of them.
– end
