Rio de Janeiro's Rio-3.5-Open-397B matches Qwen 3.7 Plus as open-source giant
A 397B-parameter open-source model from a city government rivals major labs.
The City of Rio de Janeiro's research team has released Rio-3.5-Open-397B, a massive 397-billion-parameter open-source language model on Hugging Face. Built as a fine-tune of Alibaba's Qwen architecture, the model is claimed to deliver performance on par with Qwen 3.7 Plus—a notable proprietary variant. This release is particularly striking because it originates from a municipal government rather than a big tech company or academic institution. The researchers behind the model, including those from the city's government, have made all weights and code publicly available, aligning with the open-source ethos.
Rio-3.5-Open-397B represents one of the largest openly released models from any government entity. With 397B parameters, it rivals models like Llama 3.1 405B in scale and could serve as a powerful base for further fine-tuning in Portuguese and specialized domains. The model's open license allows unrestricted use, including commercial applications, which could accelerate AI adoption in public services. However, users will need substantial compute resources to run inference. The release highlights a growing trend of public-sector involvement in foundational AI development, potentially influencing policy and technology sovereignty debates.
- Rio-3.5-Open-397B has 397 billion parameters and is fully open-source on Hugging Face.
- It is a fine-tuned Qwen model that matches the performance of Qwen 3.7 Plus according to the developers.
- The model was built by the Rio de Janeiro city government's research team, a rare public-sector AI breakthrough.
Why It Matters
A city government releasing a 397B open-source model could democratize AI and inspire public-sector innovation.