license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: mlx
tags:
- mlx
- qwen3_5
- image-text-to-text
- mxfp8
- apple-silicon
- uncensored
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): theqwen3_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 themtp.*weights on load. Text
and vision are unaffected; only spec-decode is not available. tie_word_embeddings=false, so the ~1.27Blm_headis 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.