Update README.md

README.md

@@ -7,4 +7,161 @@ base_model:
 - Qwen/Qwen3-235B-A22B-Thinking-2507
 pipeline_tag: text-generation
 library_name: transformers
----
+---
+
+# Rio 3.0 Open
+
+**Rio 3.0 Open** is a frontier-class reasoning model developed by [IplanRIO](https://iplanrio.rio.rj.gov.br/), the municipal IT company of Rio de Janeiro's city government. Built through distillation on top of Qwen3-235B-A22B-Thinking-2507 using reasoning traces from our to be announced Rio 3.0 model, Rio 3.0 Open achieves state-of-the-art results across mathematics, STEM, and code benchmarks — surpassing its base model by significant margins and competing with the world's best open and proprietary reasoning models.
+
+Rio 3.0 Open 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
+
+- **235B total / 22B active parameters** (Mixture-of-Experts)
+- **262,144 token context window**
+- **SwiReasoning integration** — dynamic explicit/latent reasoning switching for Pareto-superior accuracy and efficiency
+- **Distilled from Qwen3-235B-A22B-Thinking-2507** with advanced post-training optimization
+- **Multilingual** — strong performance in Portuguese, English, Chinese, and dozens of other languages
+- **MIT License** — fully open for commercial and research use
+
+## Benchmark Results
+
+### Mathematics & STEM
+
+| Model | GPQA Diamond | LiveCodeBench | Composite Math* | AIME 2025 | AIME 2026 I | HMMT 2025 I | HMMT 2025 II | BRUMO 2025 | CMIMC 2025 | SMT 2025 |
+|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| **Rio 3.0 Open** | **85.10%** | **76.00%** | **91.78%** | **96.67%** | **93.33%** | **90.00%** | **90.00%** | **95.00%** | **86.88%** | **90.57%** |
+| Rio 3.0 Open (w/o latent) | 83.20% | 76.00% | 89.84% | 95.00% | 89.17% | 85.83% | 90.83% | 92.50% | 85.00% | 90.57% |
+| Qwen3-235B-A22B-2507 (base) | 81.10% | 74.10% | 86.83% | 91.67% | 87.50% | 83.33% | 89.17% | 87.50% | 83.75% | 84.91% |
+| Kimi K2.5 Thinking | 87.60% | 85.00% | 93.12% | 95.83% | 93.33% | 93.33% | 89.17% | 98.33% | 91.25% | 90.57% |
+| DeepSeek V3.2 | 82.40% | 83.30% | 90.93% | 94.17% | 91.67% | 92.50% | 90.00% | 96.67% | 83.75% | 87.74% |
+| GLM 4.6 | 81.00% | 82.80% | 91.69% | 91.67% | 91.67% | 93.33% | 91.67% | 94.17% | 88.75% | 90.57% |
+| GPT OSS 120B | 80.10% | 77.97% | 89.17% | 90.00% | 89.17% | 90.00% | 90.00% | 91.67% | 85.62% | 87.74% |
+| GPT OSS 20B | 71.50% | 70.26% | 82.34% | 89.17% | 85.00% | 76.67% | 83.33% | 86.67% | 72.50% | 83.02% |
+
+*Composite Math is the average across all other mathematics benchmarks in this table.
+
+### Rio Model Family Comparison
+
+| Model | GPQA Diamond | LiveCodeBench | Composite Math* | AIME 2025 |
+|:---|:---:|:---:|:---:|:---:|
+| **Rio 3.0 Open** | **85.10%** | **76.00%** | **91.78%** | **96.67%** |
+| Rio 2.5 Open | 77.20% | 69.60% | 87.53% | 93.33% |
+| Rio 3.0 Open Mini | 71.90% | 63.50% | 78.11% | 89.17% |
+
+### Gains Over Base Model (Qwen3-235B-A22B-2507)
+
+| Benchmark | Base Model | Rio 3.0 Open | Δ |
+|:---|:---:|:---:|:---:|
+| GPQA Diamond | 81.10% | 85.10% | **+4.00%** |
+| LiveCodeBench | 74.10% | 76.00% | **+1.90%** |
+| Composite Math | 86.83% | 91.78% | **+4.95%** |
+| AIME 2025 | 91.67% | 96.67% | **+5.00%** |
+| AIME 2026 I | 87.50% | 93.33% | **+5.83%** |
+| HMMT 2025 I | 83.33% | 90.00% | **+6.67%** |
+| BRUMO 2025 | 87.50% | 95.00% | **+7.50%** |
+| CMIMC 2025 | 83.75% | 86.88% | **+3.13%** |
+| SMT 2025 | 84.91% | 90.57% | **+5.66%** |
+
+## SwiReasoning: Latent/Explicit Reasoning
+
+Rio 3.0 Open 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.
+
+The benchmark table above includes **(w/o latent)** rows showing performance with standard explicit-only reasoning, demonstrating the consistent gains from SwiReasoning across all benchmarks.
+
+## How to Use
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "prefeitura-rio/Rio-3.0-Open"
+
+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-3.0-Open \
+    --tensor-parallel-size 4 \
+    --max-model-len 262144 \
+    --trust-remote-code
+```
+
+### Using with SGLang
+
+```bash
+python -m sglang.launch_server \
+    --model-path prefeitura-rio/Rio-3.0-Open \
+    --tp 4 \
+    --context-length 262144 \
+    --trust-remote-code
+```
+
+## Model Details
+
+| | |
+|:---|:---|
+| **Developer** | IplanRIO — Empresa Municipal de Informática e Planejamento S.A. |
+| **Base Model** | Qwen3-235B-A22B-Thinking-2507 |
+| **Architecture** | Mixture-of-Experts (MoE) Transformer |
+| **Total Parameters** | ~235B |
+| **Active Parameters** | ~22B |
+| **Context Length** | 262,144 tokens |
+| **Default Max Output Length** | 81,920 tokens |
+| **Training Method** | Distillation + post-training optimization |
+| **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 3.0 Open is built upon the exceptional work of the [Qwen Team](https://github.com/QwenLM) and their Qwen3 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.rio.rj.gov.br/).