library_name: mlx
license: apache-2.0
pipeline_tag: image-text-to-text
base_model: mlx-community/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated-4bit
base_model_relation: quantized
tags:
- mlx
- qwen3_5
- vision
- abliterated
- 4-bit
Qwen3.5-4B-4bit-Abliterated
Standalone MLX 4-bit image-and-text model, repackaged for Dream Vault and mlx-swift-lm compatibility from mlx-community/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated-4bit.
This is the Huihui abliterated derivative of the Jackrong Claude 4.6 Opus reasoning-distilled Qwen3.5 model. These are upstream lineage labels; this release makes no claim of affiliation with Anthropic.
The model already contains its vision encoder. All safetensors are preserved byte for byte. No training or further abliteration was performed. Quantization is 4-bit affine with group size 64, as supplied upstream.
Compatibility changes
Added a flat preprocessor_config.json derived from the existing nested image processor configuration for Swift's Qwen3VL processor. The original processor_config.json and model configuration are preserved. The upstream text-only chat template was replaced with the matching pinned official Qwen3.5 image-aware template, including the tokenizer configuration copy. Image normalization, patch size, temporal patch size, merge size and pixel limits remain those of the source.
Python image inference
python -m mlx_vlm.generate --model dream-vault-community/Qwen3.5-4B-4bit-Abliterated --image /path/to/image.png --prompt "Describe this image." --max-tokens 256 --temperature 0.0
Use an Apple Silicon Mac with MLX and a version of mlx-vlm supporting Qwen3.5. Local release directories can replace the model repository argument. Runtime validation results, when available, are recorded separately; this card does not imply testing of unlisted runtimes or video inference.
Provenance and license
Pinned source revision: a0cd61406062a7bee31fc995a806fdb876f318ed. See PROVENANCE.json, SHA256SUMS and the unchanged UPSTREAM_README.md for source details. The immediate upstream declares Apache-2.0 but does not ship a license file; LICENSE is copied from the corresponding original Qwen3.5 release.
Abliteration is the upstream author's characterization, not a guarantee of any particular behavior. Generated image descriptions can be inaccurate; model quality and safety behavior were not comprehensively benchmarked for this compatibility repackaging.
Validation results
See validation/REPORT.md for runtime versions, synthetic image inference results, and limitations. The public fixture and Swift output are included in validation/. This is a smoke test, not a general image-understanding benchmark.