---
license: mit
base_model: prefeitura-rio/Rio-3.1-Open-30B
base_model_relation: quantized
pipeline_tag: text-generation
library_name: mlx
tags:
- mlx
- qwen3-moe
- moe
- text-generation
---

# Rio-3.1-Open-30B-MLX-8bit

MLX (Apple Silicon) conversion of [prefeitura-rio/Rio-3.1-Open-30B](https://huggingface.co/prefeitura-rio/Rio-3.1-Open-30B) (Qwen3-MoE, 128 experts), quantized to **8-bit**. First MLX build of this model.

## Quantizations
Part of the [**Rio-3.1-Open-30B MLX** collection](https://huggingface.co/collections/pipenetwork/rio-31-open-30b-mlx-6a30775145e6463297878a39).

| Variant | Notes |
|---|---|
| **8-bit** (this repo) | 8-bit · near-lossless |
| [6-bit](https://huggingface.co/pipenetwork/Rio-3.1-Open-30B-MLX-6bit) | 6-bit · high quality |
| [5-bit](https://huggingface.co/pipenetwork/Rio-3.1-Open-30B-MLX-5bit) | 5-bit |
| [4-bit](https://huggingface.co/pipenetwork/Rio-3.1-Open-30B-MLX-4bit) | 4-bit · balanced default |

## Use with mlx-lm
```bash
pip install mlx-lm
python -m mlx_lm generate --model pipenetwork/Rio-3.1-Open-30B-MLX-8bit --prompt "Olá, tudo bem?" -m 256
```

## Validation
Smoke-tested locally: loads and generates coherent text.

## License
MIT (inherited from base). Quantization config: `{"group_size": 64, "bits": 8, "mode": "affine", "model.layers.0.mlp.gate": {"group_size": 64, "bits": 8}, "model.layers.1.mlp.gate": {"group_size": 64, "bits": 8}, "model.layers.2.mlp.gate": {"group_s`.