Rio de Janeiro's municipal IT company, IplanRIO, has open-sourced a massive AI model on Hugging Face: Rio-3.5-Open-397B. The model is a fine-tuned version of Alibaba's Qwen 3.5 397B, using a Mixture of Experts (MoE) architecture with roughly 397 billion total parameters and about 17 billion active per token. It supports a context window of around 1 million tokens and is released under the MIT license.
The real story is the reasoning technique it employs. According to the model card, Rio-3.5-Open-397B integrates SwiReasoning, a training-free method that dynamically switches between explicit chain-of-thought reasoning and implicit vector-space reasoning based on information entropy. Explicit reasoning writes conclusions in natural language tokens, while implicit reasoning explores multiple paths in a hidden space, cutting down on unnecessary text output.
The results are striking. With implicit reasoning enabled, Rio-3.5-Open-397B scores 58.1 on SWE-Bench Pro and 89.5 on IMOAnswerBench. For comparison, the base Qwen 3.5 397B scores 50.9 and 80.9, respectively. A fine-tuned version without implicit reasoning hits 54.8 and 84.5. So implicit reasoning doesn't exactly double absolute performance, but it nearly doubles the improvement over the base model.
There's a catch. In the Hugging Face discussion threads, the team confirms these scores were achieved with SwiReasoning active. SwiReasoning requires probability-weighted continuous "soft embedding" vectors at inference time, which inference engines like llama.cpp—designed for discrete token IDs—can't fully handle yet. The team says the model is still significantly stronger than the base Qwen 3.5 397B without implicit reasoning, but unlocking its full potential will require inference frameworks that support soft embeddings.