Rio de Janeiro launches AI model rivaling ChatGPT, Claude
What's the story
In a surprising turn of events, the city government of Rio de Janeiro in Brazil has launched an advanced artificial intelligence (AI) model. Dubbed Rio 3.5 Open 397B, this 397-billion-parameter AI is already outperforming some of the top large language models (LLMs) in the world. The development comes at a time when tech giants and well-funded start-ups are leading the AI race.
Model development
Developed by municipal IT company IplanRIO
The innovative AI model was developed by IplanRIO, a municipal IT company that manages the digital infrastructure and public services of Rio de Janeiro. The city government has made the model available on Hugging Face under an MIT license. What makes this release even more impressive is its performance in benchmarks, where it has outperformed several top open-source models such as DeepSeek V4 and even some closed models like older versions of ChatGPT and Claude.
Model architecture
Built on Alibaba's Qwen 3.5-397B model
Rio 3.5 Open 397B is built on Alibaba's Qwen 3.5-397B-A17B, one of China's most powerful open-source foundation models. Instead of training a completely new model from scratch, IplanRIO fine-tuned an existing Qwen model and added its own optimizations along with a reasoning framework called SwiReasoning. The city government's unique approach to AI development comes in stark contrast to the recent US ban on foreigners using top Anthropic AI models such as Mythos and Fable 5.
Future implications
Potential shift in AI development landscape
The launch of Rio 3.5 Open 397B could be a glimpse into the future of AI development. Training a frontier AI model from scratch requires massive amounts of money, computing power, and talent, resources that most governments, enterprises, and institutions can't afford. However, fine-tuning an already capable open-source model is a more realistic option. If successful, this approach could prompt more governments and organizations to follow suit.