---
license: apache-2.0
base_model: mlx-community/Qwen3-VL-4B-Instruct-3bit
pipeline_tag: image-text-to-text
tags:
  - mlx
  - vision
---

# Qwen3-VL-4B-Instruct-3bit (repaired)

[![Open in MLXHub](https://mlxhub.app/assets/badge-open-in-mlxhub.svg)](https://mlxhub.app/open-model?repo=DreamFoundries/Qwen3-VL-4B-Instruct-3bit)

Re-upload of [mlx-community/Qwen3-VL-4B-Instruct-3bit](https://huggingface.co/mlx-community/Qwen3-VL-4B-Instruct-3bit) with one fix: the upstream repo ships a stale `model.safetensors.index.json` describing a 2-shard bf16 checkpoint (`model-00001-of-00002.safetensors` / `model-00002-of-00002.safetensors`) that does not exist in that repo — only a single consolidated `model.safetensors` (3-bit, MLX format, 1219 tensors) is actually present. Any loader that trusts the index blindly (mlx-swift-lm's `safetensorWeightURLs`, among others) fails trying to open a shard that was never there, and no re-download fixes it since the missing files simply don't exist upstream.

This repo removes the stale index. Weights are otherwise byte-identical to upstream — same single `model.safetensors`, same digest.

All credit for the conversion to MLX and the 3-bit quantization goes to [mlx-community](https://huggingface.co/mlx-community). All credit for the base model goes to the [Qwen team](https://huggingface.co/Qwen) (Apache-2.0).

## Usage

```bash
pip install -U mlx-vlm
python -m mlx_vlm.generate --model DreamFoundries/Qwen3-VL-4B-Instruct-3bit --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>
```

Or open directly in [MLXHub](https://mlxhub.app) on iOS/iPadOS via the badge above.

---

[![Download MLXHub](https://mlxhub.app/assets/badge-download-mlxhub.svg)](https://apps.apple.com/app/apple-store/id6766485144?pt=121945436&ct=HuggingFace&mt=8)