Qwen3.5-122B-A10B-abliterated-4bit-vlm-mlx-cs2764-pass2

4-bit MLX VLM release of a Qwen3.5-122B-A10B abliterated checkpoint, with an additional broad cs2764/mlx-abliteration pass applied on the MLX model.

A newer, more speed-balanced variant is available here:

• vanch007/Qwen3.5-122B-A10B-abliterated-4bit-vlm-mlx-cs2764-final

Summary

• Architecture: qwen3_5_moe / Qwen3_5MoeForConditionalGeneration
• Modality: vision-language model
• Quantization: 4-bit MLX (group_size=64, mode=affine)
• Model size on disk: about 65.1 GB
• Weight shards: 14
• Toolkit used for the extra ablation pass: cs2764/mlx-abliteration

When To Use This Repo vs The Newer final Repo

Aspect	This pass2 repo	Newer final repo
Extra ablation scope	broader pass over layers 0..47	targeted hot-path pass on layers 36..47
Goal	maximize refusal reduction	better speed / behavior tradeoff
Local behavior spot checks	stronger refusal weakening	milder than pass2, but still weaker refusal than direct 4-bit base
Local short generation speed	about 12.6 tok/s	about 24.0 tok/s
Recommendation:

• Use this pass2 repo if you want the more aggressive refusal-removal behavior.
• Use the newer final repo if you want a faster model with a narrower targeted extra pass.

Lineage

The release chain for this repository is:

1. Foundation model: Qwen/Qwen3.5-122B-A10B
2. Input checkpoint: a user-supplied abliterated Qwen3.5-122B-A10B VLM checkpoint
3. MLX VLM conversion: quantized to 4-bit MLX while preserving vision_config, preprocessor_config.json, processor_config.json, and video_preprocessor_config.json
4. Extra MLX abliteration pass: run with cs2764/mlx-abliteration on the converted MLX model

What Is In This Repo

This repository contains a complete MLX VLM checkpoint:

• config.json with vision_config
• model.safetensors.index.json
• model-00001-of-00014.safetensors through model-00014-of-00014.safetensors
• tokenizer.json, tokenizer_config.json, vocab.json
• preprocessor_config.json, processor_config.json, video_preprocessor_config.json
• abliteration_log.json for the pass-2 MLX abliteration run

Note that ablation_meta.json is inherited from the input checkpoint. For the extra MLX pass in this repository, use abliteration_log.json and this model card as the authoritative record.

Pass-2 Abliteration Configuration

The additional MLX abliteration pass was run with the following settings:

Parameter	Value
Toolkit	cs2764/mlx-abliteration
Refusal vector policy	per-layer
Ablation vector source	per-layer
Ablation strength	2.0
Refusal direction method	projected
Probed layers	0..47
Adaptive search	False
Attention only	False
MoE safe mode	True
Timestamp	2026-03-07T02:40:35Z

This run was done on the MLX model, not on a raw Transformers checkpoint.

Compatibility

This repository is for MLX / Apple Silicon usage. It is not a standard Transformers-only release.

Verified locally:

• mlx_vlm.load(..., lazy=True) loads successfully
• processor_class = Qwen3VLProcessor
• has_vision_tower = True
• config.json retains vision_config

Usage

Python with mlx-vlm

from mlx_vlm import load, generate

model, processor = load(
    "vanch007/Qwen3.5-122B-A10B-abliterated-4bit-vlm-mlx-cs2764-pass2",
    lazy=True,
)

result = generate(
    model,
    processor,
    prompt="Describe the image briefly.",
    image="/absolute/path/to/example.jpg",
    max_tokens=128,
    verbose=False,
)

print(result.text)

Safety Notice

This repository contains a model with reduced refusal behavior. It may produce harmful, offensive, or unsafe content.

Do not use it in consumer-facing or safety-sensitive systems without independent safety controls.

You are responsible for ensuring compliance with applicable law, policy, and platform rules.