← back to catalog · registered 2026-09-19 13:56

prithivMLmods/Qwen3.8-27B-abliterated-MLX

prithivMLmods 27B multimodal second-order
curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/prithivMLmods%2FQwen3.8-27B-abliterated-MLX"
Response includes
  • classification m1
  • files 23
  • author_summary 97 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
0
Likes
1
Model age
today
created 2026-09-19

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 25K downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
apache-2.0
Tags
mlx safetensors qwen3_5 text-generation-inference mlx-vlm uncensored abliterated apple-silicon 4-bit 8-bit image-text-to-text conversational

Related

Total size
51.0 GB
Files
23
Quantizations
1
Registered
2026-09-19 13:56
Last updated on HF
2026-09-19 13:39

Files by quantization

Auxiliary files 23 files 51.0 GB
model-00009-of-00011.safetensors 4.98 GB 0cebdf31 download
model-00005-of-00011.safetensors 4.98 GB 3956f116 download
model-00007-of-00011.safetensors 4.96 GB 4bdb1555 download
model-00004-of-00011.safetensors 4.96 GB 7f9c03df download
model-00006-of-00011.safetensors 4.96 GB fffa66bb download
model-00008-of-00011.safetensors 4.96 GB a6ac12db download
model-00003-of-00011.safetensors 4.96 GB 50afb376 download
model-00002-of-00011.safetensors 4.96 GB 48e0542e download
model-00001-of-00011.safetensors 4.87 GB 6750a23e download
model-00010-of-00011.safetensors 4.03 GB f9c6c237 download
model-00011-of-00011.safetensors 2.37 GB 1219c9ea download
tokenizer.json 19.1 MB 06b95093 download
vocab.json 6.41 MB 0aa0ce06 download
model.safetensors.index.json 113 KB 3425801b download
chat_template.jinja 8.74 KB c0c686f9 download
README.md 5.19 KB e3798f9f download
config.json 4.62 KB 5abfcb09 download
.gitattributes 1.64 KB 3539bbc1 download
tokenizer_config.json 1.14 KB 1d134cd2 download
processor_config.json 991 B 8f29fe38 download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


language: en
library_name: mlx
base_model_relation: quantized
pipeline_tag: image-text-to-text
tags:

  • mlx
  • text-generation-inference
  • qwen3_5
  • mlx-vlm
  • uncensored
  • abliterated
  • apple-silicon
  • 4-bit
  • 8-bit
    license: apache-2.0
    base_model:
  • huihui-ai/Huihui-Qwen3.8-27B-abliterated

Qwen3.8-27B-abliterated-MLX

Qwen3.8-27B-abliterated is an uncensored version of Qwen/Qwen3.8-27B created by huihui-ai through abliteration — a crude, proof-of-concept activation-editing technique that removes refusal behavior directly from model weights without relying on TransformerLens. In its latest update, only layers 18 through 51 are ablated (previously the first 15 layers were also retained unablated, and remain so), leaving the earliest layers, MTP, and visual components unmodified in an effort to preserve more of the base model's original performance; a prior revision (d42ca89) is available for those who want the earlier ablation range. The model is usable directly via Hugging Face Transformers (with enable_thinking togglable at inference) or through Ollama via huihui_ai/Qwen3.8-abliterated, and ships with explicit usage warnings: it has undergone no rigorous safety optimization, may produce sensitive or inappropriate content, is recommended only for research or controlled testing rather than public-facing production use, and places full legal and ethical responsibility for outputs on the user, with huihui.ai disclaiming liability. It's released under the Apache 2.0 license.

Model Variants & Storage Footprint

[!NOTE]
View the tree here: huggingface-tree.hf.space

prithivMLmods/Qwen3.8-27B-abliterated-MLX (main)  [~93.83 GB Total]
├── 4bit/                                        [~15.03 GB]
│   ├── model shards: 3 safetensors (~15.00 GB)
│   └── metadata & configs (~26.7 MB)
│
├── 8bit/                                        [~27.53 GB]
│   ├── model shards: 6 safetensors (~27.50 GB)
│   └── metadata & configs (~26.7 MB)
│
└── [BF16 Base Files]                            [~51.33 GB]
    ├── model shards: 11 safetensors (~51.30 GB)
    └── metadata & configs (~25.5 MB)

Use with mlx

pip install -U mlx-vlm

BF16 Variant (Base Model)

Because the BF16 weights and configurations are located in the repository root, reference the model repository directly:

python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.8-27B-abliterated-MLX \
  --max-tokens 100 \
  --temperature 0.0 \
  --prompt "Describe this image." \
  --image <path_to_image>

8-bit Variant

Load the 8-bit quantized weights from the 8bit subfolder:

CLI (Command Line)

python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.8-27B-abliterated-MLX/8bit \
  --max-tokens 100 \
  --temperature 0.0 \
  --prompt "Describe this image." \
  --image <path_to_image>

Python API

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

# Load model from the 8bit subfolder
model_path = "prithivMLmods/Qwen3.8-27B-abliterated-MLX"
model, processor = load(model_path, subfolder="8bit")
config = load_config(model_path, subfolder="8bit")

image = ["<path_to_image>"]
prompt = "Describe this image."
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=len(image))

output = generate(model, processor, formatted_prompt, image=image, max_tokens=100, temperature=0.0)
print(output.text)

4-bit Variant

Load the 4-bit quantized weights from the 4bit subfolder:

CLI (Command Line)

python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.8-27B-abliterated-MLX/4bit \
  --max-tokens 100 \
  --temperature 0.0 \
  --prompt "Describe this image." \
  --image <path_to_image>

Python API

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

# Load model from the 4bit subfolder
model_path = "prithivMLmods/Qwen3.8-27B-abliterated-MLX"
model, processor = load(model_path, subfolder="4bit")
config = load_config(model_path, subfolder="4bit")

image = ["<path_to_image>"]
prompt = "Describe this image."
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=len(image))

output = generate(model, processor, formatted_prompt, image=image, max_tokens=100, temperature=0.0)
print(output.text)

License and Attribution

This model is based on and/or incorporates the following open-source projects and models:

This model is released under the Apache License 2.0.

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in app" button that hands off directly to a local runtime of your choice - Infrahuman, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.