---
library_name: transformers
language:
- pt
- en
license: mit
base_model:
- Qwen/Qwen3.5-397B-A17B
pipeline_tag: image-text-to-text
---

# Rio 3.5 Open 397B

![Rio 3.5 Open 397B benchmark results](rio-3.5-open-benchmarks.png)

**Rio 3.5 Open 397B** is a frontier-class general-purpose AI model developed by [IplanRIO](https://iplanrio.rio.rj.gov.br/), the municipal IT company of Rio de Janeiro's city government. Post-trained from Qwen 3.5 397B, Rio 3.5 Open 397B delivers state-of-the-art open-model performance across agentic coding, mathematics, STEM, multilingual, and multimodal benchmarks — surpassing its base model by significant margins and competing with the world's best open and proprietary models.

The model is built via a merge of https://huggingface.co/nex-agi/Nex-N2-Pro and https://huggingface.co/Qwen/Qwen3.5-397B-A17B, proceeded by On-Policy Distillation from a stronger model. We detected an incorrect upload in the previous version, where the base merged version was upload instead of the final distilled model. We are sorry for the confusion and apologize profusely.

Rio 3.5 Open 397B features **SwiReasoning**, a training-free inference framework based on [Shi et al. (2025)](https://arxiv.org/abs/2510.05069) that dynamically switches between explicit chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals. This enables both higher accuracy and dramatically improved token efficiency. This model was explicitly trained to maximize the efficiency gained via latent reasoning.

## Key Features

- **397B total / 17B active parameters** (Mixture-of-Experts)
- **1,010,000 token (1M) context window**
- **SwiReasoning integration** — dynamic explicit/latent reasoning switching for Pareto-superior accuracy and efficiency
- **General-purpose** — strong agentic coding, reasoning, instruction-following, and multimodal performance
- **Post-trained from Qwen 3.5 397B**
- **Multilingual** — strong performance in Portuguese, English, Chinese, and dozens of other languages
- **MIT License** — fully open for commercial and research use

## Benchmark Results

### Agentic Coding & Software Engineering

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| Terminal-Bench 2.1 | 70.8 | 52.5 | 70.3 | 67.9 | 66.7 | **78.2** |
| DeepSWE | 23.0 | 6.0 | – | 8.0 | 24.0 | **70.0** |
| SWE-Bench Pro | 58.1 | 50.9 | 57.6 | 59.0 | **59.5** | 58.6 |
| SWE-Bench Verified | 80.2 | 76.2 | 77.7 | 80.6 | 80.2 | **82.9** |
| SWE-Bench Multilingual | **77.0** | 69.3 | 75.8 | 76.2 | 76.7 | – |

### Knowledge & Reasoning

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| GPQA Diamond | 90.9 | 88.4 | 90.3 | 90.1 | 90.5 | **93.6** |
| HLE | 36.5 | 28.7 | 34.7 | 37.7 | 36.4 | **41.4** |
| MMLU-Pro | 88.0 | 87.8 | **88.5** | 87.5 | 87.1 | – |
| MMLU-Redux | 94.6 | 94.9 | 94.5 | 94.8 | **95.3** | – |
| SuperGPQA | **72.3** | 70.4 | 71.4 | 69.9 | 71.3 | – |
| Apex | 29.2 | 9.4 | 22.7 | 38.3 | 24.0 | **80.2** |

### Mathematics

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| HMMT 2026 Feb | 93.9 | 87.9 | 92.9 | 95.2 | 92.7 | **98.5** |
| IMOAnswerBench | 89.5 | 80.9 | 86.0 | **89.8** | 86.0 | – |

### Multilingual

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| MMMLU | **89.8** | 88.5 | 89.0 | 87.9 | 87.5 | – |
| MMLU-ProX | **85.6** | 84.7 | 85.4 | 83.9 | 83.7 | – |

### Multimodal

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| MMMU-Pro | 78.4 | 79.0 | 79.0 | – | 79.4 | **81.2** |
| MathVision | 89.1 | 88.6 | **90.3** | – | 87.4 | – |
| VideoMMMU | 81.6 | 84.7 | 85.4 | – | – | **86.4** |

### Agents & Instruction Following

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| MCP-Atlas | 74.2 | 74.2 | 73.2 | 73.6 | 66.6 | **75.3** |
| IFBench | 78.4 | 76.5 | **79.1** | 77.0 | 76.0 | 76.0 |
| IFEval | 93.4 | 92.6 | **94.6** | 91.9 | 94.5 | – |

### Economic Value

| Benchmark | Rio 3.5 Open 397B | Qwen 3.5 397B (base) | Qwen 3.7 Plus | DeepSeek V4 Pro | Kimi-K2.6 | GPT 5.5 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|
| GDPval (estimated) | 1533 | 1200 | 1520 | 1554 | 1482 | **1769** |

### Gains Over Base Model (Qwen 3.5 397B)

| Benchmark | Base Model | Rio 3.5 Open 397B | Δ |
|:---|:---:|:---:|:---:|
| Terminal-Bench 2.1 | 52.5 | 70.8 | **+18.3** |
| DeepSWE | 6.0 | 23.0 | **+17.0** |
| SWE-Bench Pro | 50.9 | 58.1 | **+7.2** |
| SWE-Bench Verified | 76.2 | 80.2 | **+4.0** |
| SWE-Bench Multilingual | 69.3 | 77.0 | **+7.7** |
| GPQA Diamond | 88.4 | 90.9 | **+2.5** |
| HLE | 28.7 | 36.5 | **+7.8** |
| HMMT 2026 Feb | 87.9 | 93.9 | **+6.0** |
| IMOAnswerBench | 80.9 | 89.5 | **+8.6** |
| Apex | 9.4 | 29.2 | **+19.8** |
| GDPval (estimated) | 1200 | 1533 | **+333** |

## SwiReasoning: Latent/Explicit Reasoning

Rio 3.5 Open 397B integrates [SwiReasoning](https://arxiv.org/abs/2510.05069) (Shi et al., 2025), a training-free inference framework that dynamically alternates between two reasoning modes:

- **Explicit reasoning** — standard chain-of-thought in natural language, where the model commits tokens to a single reasoning path
- **Latent reasoning** — continuous reasoning in hidden space, where the model explores multiple implicit paths simultaneously without emitting tokens

The switching is governed by **block-wise confidence** estimated from entropy trends in the next-token distribution. When confidence is low (entropy trending upward), the model enters latent mode to explore alternatives. When confidence recovers, it switches back to explicit mode to commit to a solution.

This approach achieves a **Pareto-superior** trade-off: higher accuracy at unlimited budgets *and* dramatically better token efficiency under constrained budgets. As with previous Rio generations, the model was post-trained to maximize the gains obtained from latent reasoning.

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prefeitura-rio/Rio-3.5-Open-397B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Write a poem about Rio de Janeiro."

messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=81920,
    temperature=0.6,
    top_p=0.95,
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
print(response)
```

### Using with vLLM

```bash
vllm serve prefeitura-rio/Rio-3.5-Open-397B \
    --tensor-parallel-size 8 \
    --max-model-len 1048576 \
    --trust-remote-code
```

### Using with SGLang

```bash
python -m sglang.launch_server \
    --model-path prefeitura-rio/Rio-3.5-Open-397B \
    --tp 8 \
    --context-length 1048576 \
    --trust-remote-code
```

## Model Details

| | |
|:---|:---|
| **Developer** | IplanRIO — Empresa Municipal de Informática e Planejamento S.A. |
| **Base Model** | Qwen 3.5 397B |
| **Architecture** | Mixture-of-Experts (MoE) Transformer |
| **Total Parameters** | ~397B |
| **Active Parameters** | ~17B |
| **Context Length** | 1,010,000 tokens (1M) |
| **Training Method** | Post-training |
| **Inference Enhancement** | SwiReasoning (latent/explicit switching) |
| **License** | MIT |
| **Languages** | Multilingual (en, pt, zh, ja, ko, fr, de, es, ar, and more) |

## Citation

If you use SwiReasoning, please also cite:

```bibtex
@misc{shi2025swireasoning,
    title={SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs},
    author={Dachuan Shi et al.},
    year={2025},
    eprint={2510.05069},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}
```

## Acknowledgments

Rio 3.5 Open 397B is built upon the exceptional work of the [Qwen Team](https://github.com/QwenLM) and their Qwen 3.5 model family. We also acknowledge the authors of [SwiReasoning](https://github.com/sdc17/SwiReasoning) for their innovative inference framework.

Developed in Rio de Janeiro 🇧🇷 by [IplanRIO](https://iplanrio.prefeitura.rio/).