---
language: en
library_name: mlx
base_model_relation: quantized
pipeline_tag: text-generation
tags:
- mlx
- text-generation-inference
- qwen3_5
- mlx-vlm
- prompt-rewriter
- prompt enhancer
- apple-silicon
- 4-bit
- 8-bit
license: other
license_name: qwen-research
license_link: https://huggingface.co/Qwen/Qwen-Image-2.1-PE-I2I/blob/main/LICENSE
base_model:
- Qwen/Qwen-Image-2.1-PE-T2I
---

# **Qwen-Image-2.1-PE-T2I-MLX**

> **Qwen-Image-2.1-PE-T2I** is the text-to-image prompt-rewriting model for Qwen-Image-2.1, a fine-tuned Qwen3.5-VL-9B that turns a brief image request in any language into a detailed, expanded English prompt paired with a recommended aspect ratio, acting as the bridge between casual multilingual user requests (e.g., a short Chinese phrase describing a corgi playing guitar in the rain) and the precise, richly-detailed instructions the downstream 7B-parameter DiT-based Qwen-Image-2.1 generator needs for best results. After a `<think>` reasoning block, it outputs a structured JSON object containing the `rewritten_prompt` and a `wh_ratio` value (e.g., `16:9`), which maps directly to standard render resolutions when passed into the Diffusers `QwenImage21Pipeline`. It's usable via standard Hugging Face `AutoModelForCausalLM` loading, is intended purely as a preprocessing step ahead of image generation rather than a general-purpose chat model, and is released under the Qwen Research License Agreement.

> [!NOTE]
System Prompt — https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I/blob/main/system_prompt.txt

```text
prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX (main)  [~32.82 GB Total]
├── 4bit/                                       [~5.58 GB]
│   ├── model shards: 2 safetensors (~5.56 GB)
│   └── metadata & configs (~19.1 MB)
│
├── 8bit/                                       [~9.72 GB]
│   ├── model shards: 2 safetensors (~9.70 GB)
│   └── metadata & configs (~19.1 MB)
│
└── [BF16 Base Files]                           [~17.52 GB]
    ├── model shards: 4 safetensors (~17.50 GB)
    └── metadata & configs (~19.1 MB)
```

## Use with mlx

Install the required library:

```bash
pip install -U mlx-vlm
```

> **Task Note:** `Qwen-Image-2.1-PE-T2I` is a Text-to-Image prompt enhancer/rewriter (based on Qwen VL). It takes a short, simple user prompt (e.g., *"A cybernetic tiger in Tokyo"*) and outputs an expanded, high-detail English diffusion prompt along with the optimal aspect ratio (`wh_ratio`).

### BF16 Variant (Base Weights)

The BF16 weights reside directly in the root repository path:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX"
model, processor = load(model_path)
config = load_config(model_path)

prompt = "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(model, processor, formatted_prompt, max_tokens=512, temperature=0.2)
print(output.text)
```

### 8-bit Variant

Target the `8bit` subfolder:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX/8bit \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX"
model, processor = load(model_path, subfolder="8bit")
config = load_config(model_path, subfolder="8bit")

prompt = "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(model, processor, formatted_prompt, max_tokens=512, temperature=0.2)
print(output.text)
```

### 4-bit Variant

Target the `4bit` subfolder:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX/4bit \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/Qwen-Image-2.1-PE-T2I-MLX"
model, processor = load(model_path, subfolder="4bit")
config = load_config(model_path, subfolder="4bit")

prompt = "A neon cybernetic samurai standing under rainy Tokyo streets, cinematic lighting"
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(model, processor, formatted_prompt, max_tokens=512, temperature=0.2)
print(output.text)
```

## License and Attribution

* **Base Model:** [Qwen/Qwen-Image-2.1-PE-T2I](https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I)
  * **License:** [Qwen Research License](https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I/blob/main/LICENSE)
* **MLX-VLM:** [Blaizzy/mlx-vlm](https://github.com/Blaizzy/mlx-vlm)
* **MLX:** [ml-explore/mlx](https://github.com/ml-explore/mlx)