sahilchachra/Huihui-Qwen3.8-27B-abliterated-MXFP8

Huihui-Qwen3.8-27B-abliterated — MLX MXFP8

MLX MXFP8 quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated, uncensored via abliteration (refusal-direction removal on text layers 18–51; the README notes MTP and the vision tower were left unmodified) of Qwen3.8-27B. Qwen3.8-27B is a qwen3_5 vision-language model with a hybrid GatedDeltaNet linear-attention + full-attention text backbone (64 layers, full attention every 4th) and a Qwen3-VL vision tower. Runs on Apple Silicon via mlx-vlm. Stays image-text-to-text — the vision tower is kept in bf16; only the text backbone is quantized.

Precision	MXFP8 (E4M3 + E8M0 shared scale, group size 32)
Bits per weight	8.381 bpw
On-disk size	27 GB (27 shards)
Quantized	text backbone (incl. the ~1.27B lm_head)
Kept in bf16	Qwen3-VL vision tower

Quantizations

Variant	Bits	Size
Huihui-Qwen3.8-27B-abliterated-MXFP4	4.449 bpw	14 GB	smaller / for 16 GB+
Huihui-Qwen3.8-27B-abliterated-MXFP8	8.381 bpw	27 GB	← this repo

Verification

This higher-fidelity build was verified structurally (correct tensor shapes, format: mlx metadata, consistent shard index, vision tower intact in bf16). Full token-by-token generation was not benchmarked on the 24 GB test machine because 27 GB exceeds its RAM; on a 32 GB+ Mac it runs at normal speed. Since MXFP8 uses more bits than the MXFP4 build — which passed text + vision smoke tests end-to-end — it is at least as faithful to the base model.

See the MXFP4 build for the full generation/vision smoke-test results.

Usage (mlx-vlm)

pip install -U mlx-vlm   # needs the qwen3_5 architecture (>= 0.6.12)

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("sahilchachra/Huihui-Qwen3.8-27B-abliterated-MXFP8")
config = model.config

prompt = apply_chat_template(processor, config, "What is the capital of France?")
print(generate(model, processor, prompt, max_tokens=256, verbose=True))

This is a reasoning model; it produces a <think> channel before its answer.

Run in LM Studio

Loads and runs in LM Studio (tested on 0.4.20, mlx-llm runtime): the qwen3_5 architecture is recognized, the model indexes cleanly (format: mlx present), and the ChatML template runs as-is. This is a reasoning model — it emits a thinking channel (reasoning_content) before the final content, so give it enough max_tokens (e.g. 200+) or the answer can be empty while it is still thinking. Verified: "capital of France" → reasoning + content = "Paris".

Note: at 27 GB this needs a 32 GB+ Mac to load under LM Studio's default guardrails.

Notes & limitations

• Uncensored model. This is a deliberately uncensored/abliterated derivative and will not refuse requests the way the original might. Use responsibly and in line with the base model's license and your local laws.
• MTP head dropped. The base model's multi-token-prediction (speculative decoding) head is not included — mlx-vlm strips the mtp.* weights on load. Text and vision are unaffected; only spec-decode is not available.
• tie_word_embeddings=false, so the ~1.27B lm_head is a separate matrix and is quantized. Verified correct on the MXFP4 build above.
• Inherits all capabilities and limitations of the base model. See the original model card.
• Quantized by @sahilchachra with MLX. Apache-2.0.