---
license: apache-2.0
base_model: microsoft/AesCode-32B
tags:
- mlx
- mlx-vlm
- qwen3_vl
- 4-bit
- image-text-to-text
- code-generation
- html
- design
library_name: mlx-vlm
pipeline_tag: image-text-to-text
language:
- en
---

# AesCode-32B-4bit (MLX)

MLX 4-bit quantization of [microsoft/AesCode-32B](https://huggingface.co/microsoft/AesCode-32B) for Apple Silicon.

AesCode generates information-rich visual artifacts (slides, posters, dashboards) as structured, editable HTML/CSS, using an image generated from the same prompt as an aesthetic reference. It starts from Qwen3-VL-32B-Instruct and is trained with cold-start SFT + GDPO across seven reward channels.

## Conversion

- Tool: mlx-vlm (convert), affine quantization
- Text model: uniform **4-bit, group size 64** (~4.7 GB/s avg bits 4.705 incl. bf16 vision tower)
- Vision tower: **unquantized (bf16)**
- Context: 262,144 tokens
- Size: ~19.7 GB (4 shards)
- Note: the original 0.7.1-era conversion omitted the `quantization` dict in config.json; it was reconstructed and verified. Recommended loaders: **mlx-vlm >= 0.7.6** or **oMLX >= 0.7.0**.

## Usage

```bash
pip install mlx-vlm
python -m mlx_vlm generate --model cuttie1979/AesCode-32B-4bit \
  --prompt "Describe this layout" --image reference.png --max-tokens 4096
```

Serving with [oMLX](https://github.com/jundot/omlx): place the checkpoint in your model directory and restart, then call the OpenAI-compatible `/v1/chat/completions`.

## Reference images

AesCode was trained with an optional reference image for layout/style guidance. Quality is highest with a reference, but the model degrades gracefully without one (−1.0 Visual point measured on AesCode-8B vs −19.6 for the backbone).

## License

Apache 2.0, following [microsoft/AesCode-32B](https://huggingface.co/microsoft/AesCode-32B). See the upstream model card for the training details and citation (AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards, under review 2027).