Rio de Janeiro's city government releases Rio 3.5 Open 397B: a 403-billion-parameter open model built on Qwen

The City of Rio de Janeiro has released Rio 3.5 Open 397B, a 403-billion-parameter open-weights AI model built by IplanRIO, the municipal technology company that manages the city's digital infrastructure. The model is published on Hugging Face under an MIT license and supports both Portuguese and English.

What's new

Technical specifications:

• Parameters: 403B total (mixture-of-experts architecture derived from Qwen3.5-397B-A17B)
• Architecture: built via a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by on-policy distillation from a stronger model
• SwiReasoning: a training-free inference framework that dynamically switches between explicit chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals
• Modalities: image and text input, text output
• License: MIT (fully open)
• Deployment: compatible with Transformers, vLLM, and SGLang; OpenAI-compatible API available via vLLM

The model card reports benchmark scores exceeding Qwen3.7 Plus on four of five tasks:

These numbers are first-party and have not been independently audited.

The team published an initial correction: an earlier upload contained the base merged version rather than the intended final distilled model. The corrected model is now live on Hugging Face.

Context

IplanRIO manages digital services for a city of over 6 million residents, including emergency response systems, public service automation, and smart city infrastructure. Rio de Janeiro has been investing in AI-driven public services, and this release extends that work to foundation model development.

The base model, Alibaba's Qwen3.5-397B-A17B, is one of the strongest publicly available mixture-of-experts architectures. Rio's training approach — merging with Nex-N2-Pro and applying on-policy distillation — represents a mid-tier investment rather than ground-up pretraining, making it achievable by a well-resourced municipal technology team.

Brazil has national AI programs underway, but a city-level release at this parameter scale is without clear precedent globally.

Why it matters

A municipal government releasing a competitive open frontier model is genuinely novel. Cities typically license commercial models or use smaller specialized systems. Rio 3.5 sits at the scale of the strongest open-weights models globally — if the benchmark numbers hold up under independent evaluation.

The MIT license and Hugging Face publication make the model freely usable by other governments, research institutions, and developers worldwide. The strong Portuguese language support is particularly relevant for Brazil's 210 million residents and the broader Portuguese-speaking world of roughly 270 million people.

The caveats are real: the benchmark claims are self-reported and the training methodology builds on third-party checkpoints. Independent evaluation is still pending. But the existence of the release itself — a functioning 403B open model from a city government — is the story.