divinetribe/AesCode-32B-MLX-8bit

AesCode-32B MLX 8bit

An 8-bit MLX quantization of microsoft/AesCode-32B for Apple Silicon, with the vision tower kept in bf16 so it still sees.

AesCode-32B is a Qwen3VLForConditionalGeneration finetune of Qwen/Qwen3-VL-32B-Instruct that takes a screenshot and writes the HTML and CSS to recreate it (slides, posters, dashboards, product pages). Only the language tower is quantized to 8-bit; the 351 vision_tower.* tensors stay in bfloat16, so the image path keeps its quality.

8-bit is about half the size for smaller machines; bf16 is full quality.

Sizes in this family

• AesCode-32B-MLX-bf16 (full quality)
• AesCode-32B-MLX-8bit (half size)
• AesCode-32B-MLX-4bit (quarter size)

What is in this build

precision	8-bit language tower, bf16 vision tower
vision	all 351 vision_tower.* tensors kept, vision_config and processor configs intact
size	about 34 GB on disk
framework	runs with mlx-vlm on Apple Silicon

Vision was tested before upload

Every size of this model is put through the same gate before it ships: a real screenshot of a live ineedhemp.com product page is fed in and the model is asked for an SEO, marketing and accessibility audit plus image alt-text. A smaller build is only published if it still reads the page about as well as the full-precision one. This 8-bit build passed. From its audit of the real page:

SEO: The main H1 is strong and includes the product name "Divine Crossing v5 Rebuildable Concentrate Heater". Marketing: the "Version 5 Ceramic Heater" callout, the star rating, the trust badges (shipping, warranty) and the "Add to Cart" button all read clearly. Alt-text: Main product image: "Divine Crossing v5 rebuildable concentrate heater in metallic finish with textured grip and glass chamber."

It named the real on-page elements and wrote usable alt-text, so the quantization did not break its vision.

Use

pip install mlx-vlm
python -m mlx_vlm.generate --model divinetribe/AesCode-32B-MLX-8bit \
  --image screenshot.png \
  --prompt "Recreate this screenshot as a single HTML file with inline CSS." \
  --max-tokens 1200

Runs on an M-series Mac. No cloud inference, no API keys, no telemetry.

Benchmarks

The most authoritative evaluation that actually covers this model is the one in AesCode's own paper, "AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards" (arXiv:2610.18860). There is no independent public leaderboard that lists AesCode specifically, so these numbers come from the authors' benchmark of 300 infographic-style samples, scored for rule-following correctness and visual quality. Scores quoted from the model card:

Model (with reference image)	Overall	Visual
AesCode-32B	85.89	77.44
Claude Opus 4.8 (no reference)	76.78
GPT-5.5 (no reference)	76.31
Qwen3-VL-32B base (with reference)	61.10

These are the original model's numbers. For general-vision numbers on the backbone itself, see the official Qwen3-VL evaluations (that is the base model's benchmark, not AesCode's).

Credits

This build stands entirely on other people's work. Credit where it is due:

• The AesCode team at Microsoft, who created the model and the benchmark: Xin Gao, Yifan Yang, Xin Li, Pan Wang, Yan Li, Muzhao Tian, Mingxi Cheng, Bei Liu, Kai Qiu, Qi Dai, Dong Chen, Yue Dong, Leandros Tassiulas, Rex Ying, Ji Li and Chong Luo. Paper: arXiv:2610.18860, model: microsoft/AesCode-32B.
• The Qwen team at Alibaba, for the Qwen3-VL-32B-Instruct base.
• The MLX team at Apple and Prince Canuma and the mlx-vlm contributors, for the open-source tools that let a model like this run natively and privately on a Mac.

Thank you to all of them.

Note

Quantization trades a little precision for size. It does not add knowledge and it does not make output correct. You are responsible for what you build with it.

Built and published by divinetribe · more models