prefeitura-rio/Rio-3.0-Open-Mini
Rio 3.0 Open Mini is a 4 billion parameter reasoning model developed by IplanRIO, distilled from Qwen3-4B-Thinking-2507. It features SwiReasoning, a dynamic inference framework that switches between explicit and latent reasoning, enabling superior accuracy and token efficiency. This model excels in mathematics, STEM, and code benchmarks, often outperforming larger models and its base model significantly. It supports a 262,144 token context window and is multilingual, with strong performance across many languages.
Loading preview...
Overview
Rio 3.0 Open Mini is a 4 billion parameter reasoning model developed by IplanRIO, the municipal IT company of Rio de Janeiro. It is built through distillation from Qwen3-4B-Thinking-2507, leveraging reasoning traces from the upcoming Rio 3.0 model. This model achieves strong results in mathematics, STEM, and code benchmarks, often surpassing its base model and competing with significantly larger models.
Key Features & Innovations
- SwiReasoning Integration: A unique, training-free inference framework that dynamically switches between explicit chain-of-thought and latent-space reasoning. This mechanism, guided by entropy-based confidence signals, allows for both higher accuracy and improved token efficiency.
- Distillation: Optimized from Qwen3-4B-Thinking-2507 using advanced reasoning traces.
- Extended Context Window: Features a substantial 262,144 token context window.
- Multilingual Support: Demonstrates strong performance across Portuguese, English, Chinese, and many other languages.
- MIT License: Fully open for commercial and research applications.
Performance Highlights
Rio 3.0 Open Mini shows significant gains over its base model, Qwen3-4B-2507, across various benchmarks:
- GPQA Diamond: +6.10% gain
- LiveCodeBench: +8.30% gain
- Composite Math: +6.99% gain
- HMMT 2025 I: +17.50% gain
The integration of SwiReasoning consistently improves performance, as evidenced by comparisons with the model running without latent reasoning.