Expert 3D tensor reshape and gate/up concatenation
MTP and Vision Encoder Restoration
The GGUF file does not contain Multi-Token Prediction (MTP) or vision encoder tensors — these are excluded by the llama.cpp converter that produced it. For a complete, loadable model, the following were copied verbatim from the official Qwen3.6-35B-A3B reference model:
Component
Tensors
Source
Vision encoder (model.visual.*)
333
Reference model (bit-exact copy)
MTP layers (mtp.*)
4
Reference model (bit-exact copy)
Additional vision/metadata tensors
15
Reference model (bit-exact copy)
All 352 copied tensors verified bit-exact against the reference.
Sanity Check
The recovered model was tested with vLLM (FP8 + TP2 on 2x GPUs):
Model
Harmful Coherence
Benign Coherence
Harmful Refusal
Base MoE (FP8+TP2)
100%
100%
40%
Recovered MoE (FP8+TP2)
100%
100%
0%
The recovered model achieves 100% coherence on both harmful and benign prompts, matching the base model's generation quality. The abliteration is effective: 0% refusal rate (down from the base model's 40%).
Tensor Comparison vs Base Model
Compared against the official Qwen3.6-35B-A3B base to identify abliteration modifications:
Summary
Category
Tensors
Identical to Base
Modified
GGUF-derived
693
307
386
Copied (MTP + vision)
352
352
0
Total
1045
659
386
Unchanged Tensors (identical to base)
These tensors were not modified by abliteration:
Group
Count
Note
layernorm
82
Input/post-attention layernorms
linear_attn.norm
30
Layer norms for linear attention
linear_attn.conv1d
30
Conv1d weights
linear_attn.dt_bias
30
Delta-time biases
linear_attn.A_log
30
A-log parameters
self_attn.q_norm / k_norm
22
QK norms for full attention
router_gate
41
Expert router gates
vision
333
Vision encoder
mtp
4
Multi-token prediction layers
final_norm
1
Final layer norm
Modified Tensors
Group
Total
Modified
Typical % Changed
Max Abs Diff
expert_gate_up
41
40
41–79%
1.8e-02
expert_down
41
40
42–85%
6.5e-02
shared_expert_gate
41
40
76–93%
2.5e-02
shared_expert_up
41
40
38–92%
2.1e-02
shared_expert_down
41
40
65–88%
2.4e-02
shared_expert_gate_scalar
41
16
89–99%
5.6e-03
linear_attn.out_proj
30
30
75–88%
6.5e-02
linear_attn.in_proj_qkv
30
26
73–76%
2.3e-03
linear_attn.in_proj_z
30
26
75–77%
2.0e-03
linear_attn.in_proj_a
30
26
76–78%
9.8e-04
linear_attn.in_proj_b
30
26
77–80%
9.8e-04
self_attn.o_proj
11
10
75–87%
3.2e-02
self_attn.q_proj
11
8
75–76%
1.6e-03
self_attn.k_proj
11
8
76–80%
1.2e-03
self_attn.v_proj
11
8
77–79%
2.0e-03
embed_tokens
1
1
74%
1.1e-03
lm_head
1
1
75%
1.1e-03
Key observations:
Expert and shared expert projections show the largest deviations (up to 6.5e-02 max abs diff)
Linear attention out_proj has the highest max abs diff (6.5e-02), consistent with the 27B model pattern
Router gates and normalization layers were left untouched — the abliteration targeted only projection weights
40 of 41 MoE layers have modified expert tensors; the unmodified layer's experts may have been below a threshold
Layer 0's linear attention projections are unmodified, while layers 1+ show modifications (26/30 layers affected)
This model was recovered from a lossy Q8_0 quantization. While the conversion itself is bit-exact to the GGUF source, the original quantization introduces error on the most affected tensors compared to the original BF16 weights. The abliteration modifications (up to 0.065 max abs diff) are significantly larger than the quantization noise, confirming the abliteration signal is well-preserved.
