Rio de Janeiro city releases top AI model, day after US banned Anthropic AI for foreigners

Things are moving fast in the AI world. The latest: City of Rio de Janeiro has released a fairly decent AI model that is beating some top models from world's leading AI companies.

In Short

• Rio de Janeiro city council releases a large AI model 
• The model is comparable to some top AI models in the world
• The model, Rio 3.5, is fine-tuned from Alibaba Qwen AI model

No, this is not a joke. At first glance, this might sound like one of those AI headlines that has to be fake. But it isn't. The City of Rio de Janeiro in Brazil has indeed released a fairly decent AI model, which is even beating some of top LLMs in the world.

At a time when the AI race is dominated by Big Tech and well-funded start-ups, one of the latest high-profile AI models has come from a rather unexpected place. Imagine waking up to news that the Municipal Corporation of Delhi (MCD) released a new AI model that was outperforming some versions of ChatGPT and Claude. That's more or less what has just happened in Brazil.

Rio de Janeiro's city government has released Rio 3.5 Open 397B, a 397-billion-parameter AI model that is suddenly attracting attention across the global AI community. The model was developed by IplanRIO, the municipal IT company responsible for the city's digital infrastructure and public services, and has been released on Hugging Face under an MIT licence.

More impressive than the release is the model’s score card in benchmarks. It is beating some of the top open source models, such as DeepSeek V4, and even some closed models like older versions of ChatGPT and Claude.

So, how has a city council managed to do it? Well, the 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 China's most capable open-source foundation models. In other words, IplanRIO did not train a completely new model from the ground up. Instead, it took an existing Qwen model, fine-tuned it, added its own optimisations and a reasoning framework called SwiReasoning, and then released the improved version as an open model.

A day after US banned foreigners from using top Claude

The way Rio has gone about its AI is also what makes this story so interesting. More so when seen in the context of recent US ban on foreigners using top Anthropic Claude AI models like Mythos and Fable 5. The discussion is not really about whether Rio 3.5 beats another model on a benchmark. It's about the fact that a city government was able to take an open-source foundation model and turn it into something that is now being discussed alongside some of the leading AI systems in the world.

And this could be a glimpse into the future of AI development. Training a frontier AI model from scratch requires enormous amounts of money, computing power and talent. Most governments, enterprises and institutions simply cannot afford that. Fine-tuning an already capable open-source model, however, is a much more realistic option.

If Rio's approach proves successful, we could see more governments and organisations follow the same playbook. Rather than relying entirely on American AI companies, they may choose to take powerful open-source models and adapt them for local languages, regulations and use cases.

And that possibility could worry US tech companies. While firms such as OpenAI, Anthropic and Google increasingly keep their most advanced models behind proprietary APIs, Chinese companies like Alibaba and DeepSeek have been aggressively pushing open-weight models. The more capable those models become, the easier it becomes for others to build customised AI systems on top of them.

- Ends