IplanRIO, an IT company under the Rio de Janeiro municipal government in Brazil, has unexpectedly entered the world's top tier of AI by post-training the Qwen 3.5-397B-A17B model based on Alibaba's Tongyi Qianwen 3.5-397B-A17B model and open-sourcing a large model called Rio 3.5 Open 397B.
Article author and source: Machine Heart
Today, in addition to Claude Fable 5, which is banned globally (outside the US), the AI community is also flooded with posts about an open-source model.
A Twitter user discovered that Rio 3.5 397B, an open-source model developed by an IT company under the Rio de Janeiro municipal government in Brazil, outperformed other open-source models such as Qwen 3.7 Plus in multiple benchmark tests. The underlying model of this model is Qwen3.5-397B-A17B .
Another Twitter user, "Chubby," expressed disbelief, saying, "A model I'd never heard of before is now the open-source state-of-the-art, even surpassing Qwen 3.7."
This is the second model this week to be post-trained 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 models such as GPT-5.5 and Opus 4.7.
The rise of Rio 3.5 prompted Huggingface co-founder and CEO Clem Delangue to remark, "Nothing in the future of AI is predetermined. What will happen next will give us all choices and room to act: Path 1: Closed-source APIs, centralized power, the future decided by a few in Silicon Valley and Washington. Path 2: Open-source AI, allowing everyone to participate, own, and co-build, including institutions like Rio de Janeiro. Choose your path, my friend."
Next, let's take a look at 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.
The model was post-trained on Qwen 3.5 397B and achieved leading levels among open-source models in benchmark tests such as agent programming, mathematics, STEM, multilingualism, and multimodalism. It showed significant improvement over the base model and was able to compete with the world's strongest open-source and closed-source models.
The core features of the model include the following:
- 397B total parameters, 17B activation parameters, using a hybrid expert architecture (MoE).
- 1,010,000 tokens, which is equivalent to 1 million context windows;
- Integrated SwiReasoning: Enables dynamic switching between explicit and latent space reasoning, achieving a better balance between accuracy and efficiency;
- Strong versatility: It excels in agent programming, reasoning, instruction following, and multimodal tasks;
- Strong multilingual ability: Performs well in Portuguese, English, Chinese and dozens of other languages;
- Licensed under the MIT License: Fully open and can be used for commercial and research purposes.
Among them, SwiReasoning is a training-free inference framework based on the research of Shi et al. (2025). It can dynamically switch between explicit thought chain reasoning and latent space reasoning based on entropy-based confidence signals. This can improve both accuracy and token efficiency. During training, the model is also explicitly optimized for the efficiency gains brought by latent space reasoning.
Intelligent Agent Programming and Software Engineering Benchmarking:
Knowledge and Reasoning Benchmark Test:
Mathematical benchmark tests:
Multilingual benchmarks:
Agent and instruction compliance benchmark tests:
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
