license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.5-2B-abliterated
pipeline_tag: text-generation
tags:
- mlx
- abliterated
- qwen3.5
- experimental
Minstrel Huihui Qwen3.5-2B abliterated — MLX 4-bit
Experimental, text-only MLX derivative of Huihui Qwen3.5-2B abliterated at commit b2e291a65f29a9b148981fa5299caea5d35bd4c8. The upstream card identifies Qwen3.5-2B as its base but does not specify that base revision or a complete abliteration recipe.
This package is for Minstrel's native MLX Swift helper. It removes 297 vision tensors, retains 694 text tensors, uses 4-bit affine quantization with group size 64, and caps context at 8,192 tokens. It may not load in generic Python mlx_lm builds that lack qwen3_5_text support.
Conversion used mlx-vlm==0.3.12 and mlx==0.32.3 on Apple Silicon/macOS 27.0, followed by Minstrel's derive_qwen35_text.py. minstrel-provenance.json records the pinned source, conversion tool, and hashes of the local intermediate files. The model weights are 1,059,394,783 bytes, SHA-256 67a627fad9153d3df31bf8e86449516663ed86db9e53fdfc1334435c05d042c2.
Evaluation and status
The model loads and generates through Minstrel's mlx-swift-lm 3.31.3 helper. It passed 5/7 focused Minstrel writing-operation checks, versus 4/7 for the installed Standard model on the same checks. It made a false proofreading suggestion on clean prose, returned incomplete JSON in that case, and counted a spoken line for a character who did not speak. Both models answered four ordinary romance and war-fiction prompts without refusing; these prompts do not demonstrate an improvement in false-refusal behavior. This package has not passed Minstrel's full model acceptance gates and is not selected by a production rollout.
The package includes the Apache-2.0 license from the pinned Huihui snapshot. See the upstream model card for its additional context.