TheCluster/Qwen3.5-27B-Heretic-MLX-mixed-v1

Qwen3.5-27B Heretic

Quality: quantized (mixed quants per tensor, group size: 32, 9.119 bpw)

Only the mlp layers were quantized to 4 bit, the rest are in fp16 or bf16.

Improved prefill time on older chips (M1 Max/Ultra and probably M2 gen): up to 35% faster.

This is a abliterated (uncensored) version of Qwen/Qwen3.5-27B, made using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA)

Performance

Metric	This model	Original model (Qwen/Qwen3.5-27B)
KL divergence	0.0653	0 (by definition)
Refusals	14/100	94/100

Abliteration parameters

Parameter	Value
direction_index	37.97
attn.o_proj.max_weight	1.45
attn.o_proj.max_weight_position	59.09
attn.o_proj.min_weight	1.44
attn.o_proj.min_weight_distance	34.80
mlp.down_proj.max_weight	1.43
mlp.down_proj.max_weight_position	41.91
mlp.down_proj.min_weight	0.72
mlp.down_proj.min_weight_distance	28.18

Sampling Parameters:

• I suggest using the following sets of sampling parameters depending on the mode and task type:
  • Thinking mode for general tasks:
    temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
  • Thinking mode for precise coding tasks (e.g., WebDev):
    temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
  • Instruct (or non-thinking) mode for general tasks:
    temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
  • Instruct (or non-thinking) mode for reasoning tasks:
    temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
• For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

---

Source

This model was converted to MLX format from coder3101/Qwen3.5-27B-heretic.