GLM-4.6V-Flash Heretic MLX

This is a decensored version of zai-org/GLM-4.6V-Flash, made using Heretic v1.1.0

Performance

Metric This model Original model (zai-org/GLM-4.6V-Flash)
KL divergence 0.0000 0 (by definition)
Refusals 63/100 100/100

Abliteration parameters

Parameter Value
direction_index 22.89
attn.o_proj.max_weight 1.45
attn.o_proj.max_weight_position 28.07
attn.o_proj.min_weight 1.40
attn.o_proj.min_weight_distance 13.38
mlp.down_proj.max_weight 1.19
mlp.down_proj.max_weight_position 24.88
mlp.down_proj.min_weight 0.82
mlp.down_proj.min_weight_distance 10.68

Source

This model was converted to MLX format from AiAsistent/GLM-4.6V-Flash-heretic using mlx-vlm version 0.3.12.

Downloads last month
63
Safetensors
Model size
10B params
Tensor type
F16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TheCluster/GLM-4.6V-Flash-Heretic-MLX-8bit

Quantized
(3)
this model

Collection including TheCluster/GLM-4.6V-Flash-Heretic-MLX-8bit