IplanRIO, the IT company under the municipal government of Rio de Janeiro, Brazil, has fine-tuned the Qwen 3.5-397B-A17B model from Alibaba’s Tongyi and open-sourced a large model named Rio 3.5 Open 397B, unexpectedly entering the global top tier of AI models.Article author and source: MachineHeart
Today, the AI community is also abuzz with an open-source model, in addition to the globally (non-U.S.) banned Claude Fable 5.
A Twitter blogger discovered that Rio 3.5 397B, an open-source model developed by Rio de Janeiro City Hall’s IT company, outperformed open-source models such as Qwen 3.7 Plus across multiple benchmarks, despite being based on the Qwen3.5-397B-A17B foundation model.
Another Twitter user, "Chubby," expressed disbelief: "A model I've never heard of before is now the open-source SOTA, even surpassing Qwen 3.7."
This is already the second model this week to undergo post-training based on Qwen3.5 and reach state-of-the-art performance. The other is Nex-N2-Pro 397B, also built on Qwen3.5-397B-A17B, with performance comparable to top-tier models such as GPT-5.5 and Opus 4.7.
The rise of Rio 3.5 prompted Clem Delangue, co-founder and CEO of Hugging Face, to remark, "Nothing about the future of AI is predetermined. What happens next is up to each of us—we have choices and room to act: Path one: closed-source APIs, centralized power, with the future decided by a few in Silicon Valley and Washington. Path two: open-source AI, enabling everyone to participate, own, and co-build—including institutions like the city of Rio. Choose your path, friend."
Next, let’s explore the technical details of the Rio 3.5 397B model.
Rio 3.5 Open 397B is a cutting-edge general-purpose AI model developed by IplanRIO, the municipal IT company under the Rio de Janeiro city government.
This model has been post-trained on Qwen 3.5 397B and achieves state-of-the-art performance among open-source models on benchmarks including agent programming, mathematics, STEM, multilingual, and multimodal tasks, showing significant improvements over the base model and competing with the world’s strongest open-source and proprietary models.
The model's core features include the following:
- 397B total parameters, 17B activated parameters, using a Mixture of Experts (MoE) architecture;
- 1,010,000 tokens, equivalent to a context window of approximately one million;
- Integrate SwiReasoning: Dynamically switch between explicit reasoning and latent space reasoning to achieve an optimal balance between accuracy and efficiency;
- Strong general capabilities: excels in agent programming, reasoning, instruction following, and multimodal tasks;
- Strong multilingual capabilities: Performs well in Portuguese, English, Chinese, and dozens of other languages;
- Licensed under the MIT License: Fully open source, usable for commercial and research purposes.
Among these, SwiReasoning is a training-free reasoning framework based on Shi et al.'s 2025 research that dynamically switches between explicit chain-of-thought reasoning and latent-space reasoning using entropy-based confidence signals. This approach enhances accuracy while significantly improving token efficiency. The model is also explicitly optimized during training to maximize the efficiency gains from latent-space reasoning.
Agent Programming and Software Engineering Benchmark:
Knowledge and Reasoning Benchmark:
Mathematical benchmark:
Multilingual Benchmarking:
Agent and Instruction Following Benchmark:
Reference link:
https://x.com/kimmonismus/status/2065911865390063791
https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B
https://x.com/ZenMagnets/status/2065796012820848699
https://x.com/SemiAnalysis_/status/2065894494935933191