Rio de Janeiro’s municipal IT company, IplanRIO, has released a powerful new artificial intelligence model named Rio 3.5 Open 397B. The model, available on Hugging Face under an MIT license, scored 70.8 on Terminal-Bench 2.1, placing it ahead of DeepSeek V4 Pro, which scored 67.9.
Rather than training a model from scratch, IplanRIO fine-tuned Alibaba’s open-source Qwen 3.5-397B-A17B. The architecture uses a Mixture-of-Experts design, holding 397 billion total parameters but activating only 17 billion at any time. It features a proprietary inference framework called SwiReasoning and supports a one-million-token context window.
The release is part of Rio's broader “AI City” initiative, backed by a reported $550 million investment in data centers, aiming for what officials call “technological sovereignty” from dominant Silicon Valley firms.
The launch, however, was immediately met with controversy. AI firm Nex-AGI alleged the model’s weights are essentially a 60/40 blend of their own model and the base Qwen model, lacking independent training. IplanRIO labeled the claims an “operational error,” a vague response that neither confirms nor denies the allegation. If the claim of weight merging is verified, it could seriously damage Rio’s credibility in the global AI community.