divinetribe/AesCode-32B-abliterated-MLX-bf16

AesCode-32B Abliterated MLX bf16

Refusal-direction-abliterated microsoft/AesCode-32B for Apple Silicon, full-precision bfloat16, with the vision tower intact.

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). This build removes the model's refusal direction while keeping its coding and vision ability, and keeps all 351 vision tensors so it still sees.

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

Sizes in this family

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

Each size was put through the same ineedhemp.com product-page vision gate before it shipped. For the faithful (non-abliterated) builds, see the standard line.

What was done

method	Arditi et al. (2024) directional ablation
direction from	layer 44 of 64, chosen by the strongest normalized harmful/harmless separation in the mid-stack window
calibration	32 harmful vs 32 harmless instructions, chat-formatted, last-token residual
orthogonalized	embed_tokens + every layer's attention out-projection (self_attn.o_proj) + every mlp.down_proj
vision tower	untouched; the 351 vision_tower.* tensors and vision_config are preserved
precision	bfloat16, no quantization

Only the language tower is modified. The refusal direction is projected out of every weight that writes to the residual stream, so the model stops steering toward a refusal, without touching the parts of the network that read the image.

Quality and vision were tested before upload, side by side with the faithful build

Removing a direction can quietly degrade a model, so this build was checked on the same real task as the faithful conversion: a screenshot of a live ineedhemp.com product page, with a request for an SEO, marketing and accessibility audit plus alt-text. The abliterated model read the page just as well as the faithful one.

From its audit of the real page:

SEO: The page needs a clear title tag, probably "Divine Crossing v5 Rebuildable Concentrate Heater - Divine Tribe". The H1 is strong and includes the product name and key terms like "rebuildable concentrate heater".

Marketing/conversion: The yellow outline around the product card is loud but effective. The "Add to Cart" button is bright and good, but the quantity selector is tiny. The cup replacement explanation is useful but buried.

Alt-text: Main product image: "Divine Crossing v5 rebuildable concentrate heater with ceramic cup and metal body." Thumbnail 3: "White ceramic heater cup stacked." Thumbnail 4: "Clear glass chamber with metal base."

It named the real on-page elements (product title, price range, star rating, the yellow product card, the add-to-cart button, the tiny quantity selector, the real features like the ceramic cup and vortex airflow) and wrote usable alt-text, matching the faithful build's quality side by side. The orthogonalization did not hurt its coding or its vision.

Use

pip install mlx-vlm
python -m mlx_vlm.generate --model divinetribe/AesCode-32B-abliterated-MLX-bf16 \
  --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. Abliteration targets the refusal direction only and is not expected to change coding or vision quality, which is why this build was re-tested on the product-page task above. For general-vision numbers on the backbone, 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.
• Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee and Neel Nanda, whose paper "Refusal in Language Models Is Mediated by a Single Direction" (arXiv:2406.11717) is the technique this build uses.
• 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

Abliteration removes the model's refusal direction. 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