---
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE
language:
- en
- zh
library_name: transformers
tags:
- qwen3.6
- safetensors
- uncensored
- abliterated
- gguf-recovery
- moe
base_model:
- Qwen/Qwen3.6-35B-A3B
pipeline_tag: text-generation
---

# Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

💬 **Community:** Join the [Abliterlitics Discord](https://discord.gg/AqmDnBjPvM) for discussion, model releases and support.


Recovered HuggingFace safetensors from the Q8_0 quantized GGUF published by HauhauCS.

## Source

| Field | Value |
|-------|-------|
| **Original GGUF** | `Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf` |
| **GGUF Size** | 41 GB |
| **Quantization** | Q8_0 (355 tensors), F32 (301 tensors), F16 (77 tensors) |
| **Reference Model** | `Qwen3.6-35B-A3B` (official, BF16) |
| **Architecture** | `Qwen3_5MoeForConditionalGeneration` (MoE hybrid Gated DeltaNet + Gated Attention, 256 experts with 8 active per token) |

## Recovery Details

Converted from GGUF to HuggingFace safetensors format using [ungguf](https://github.com/dreamfast/ungguf) with bit-exact verification.

**All 693 GGUF-derived tensors verified bit-exact** against the GGUF source after applying:
- GGML Fortran-order reversal (`reverse_shape=True` for all tensors)
- Norm convention (subtract 1.0)
- A_log convention (log(-A))
- V-head inverse reorder (v_per_k=2: 16 K-heads / 32 V-heads)
- 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)

## Output Format

| Property | Value |
|----------|-------|
| **Format** | HuggingFace safetensors (17 shards) |
| **Dtype** | BF16 (dequantized from Q8_0/F32/F16) |
| **Total Size** | 67 GB |
| **Tensor Count** | 1045 |
| **Shard Size** | ~4.1 GB |

## Usage

Load with HuggingFace transformers:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "./Qwen3.6-35B-A3B-HauhauCS-Q8KP-recovered",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("./Qwen3.6-35B-A3B-HauhauCS-Q8KP-recovered")
```

For efficient inference with vLLM:

```bash
vllm serve ./Qwen3.6-35B-A3B-HauhauCS-Q8KP-recovered --quantization fp8 --tensor-parallel-size 2
```

See our other tensor comparisons and provenance analyses for HauhauCS models at:
**[DreamFast HauhauCS Safetensor Benchmarks](https://huggingface.co/collections/DreamFast/hauhaucs-safetensor-benchmarks)**

## Quality Notes

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.

## Benchmarks

**Benchmarks and tensor analysis coming soon.** See our previous HauhauCS model benchmarks and evaluations at:
**[DreamFast HauhauCS Safetensor Benchmarks](https://huggingface.co/collections/DreamFast/hauhaucs-safetensor-benchmarks)**

## Files

```
Qwen3.6-35B-A3B-HauhauCS-Q8KP-recovered/
├── config.json
├── generation_config.json
├── tokenizer.json
├── tokenizer_config.json
├── preprocessor_config.json
├── video_preprocessor_config.json
├── chat_template.jinja
├── vocab.json
├── merges.txt
├── model.safetensors.index.json
├── model.safetensors-00001-of-00017.safetensors
├── ...
├── model.safetensors-00017-of-00017.safetensors
└── diff_report.json              # Full tensor-by-tensor comparison
```