Quality: quantized (mixed quants per tensor, group size: 32, 10.101 bpw). Most layers use 8-bit affine quantization with a group size 32; some layers are saved in bf16.

Fully uncensored and fine-tuned (by DavidAU) using Claude 4.6 large distill dataset.

This version is INSTRUCT, with modified jinja template which put this model into "instruct only" mode.

The model weights were updated on April 14.

Abliteration metrics

Metric	This model	Original model (Qwen/Qwen3.5-9B)
KL divergence	0.0793	0 (by definition)
Refusals	6/100	100/100

Benchmarks:

         arc   arc/e boolq hswag obkqa piqa  wino

HERETIC verison (this model):
mxfp8    0.574,0.755,0.869,0.714,0.410,0.780,0.691

Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT
mxfp8    0.574,0.729,0.882,0.711,0.422,0.775,0.691

Qwen3.5-9B
mxfp8    0.417,0.458,0.623,0.634,0.338,0.737,0.639

Source

This model was converted to MLX format from DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED using mlx-vlm version 0.4.4.

TheCluster/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mixed-10bit

Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED

Abliteration metrics

Benchmarks:

Source

TheCluster/
Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mixed-10bit