README.md

---
license: apache-2.0
tags:
  - uncensored
  - qwen3.6
  - moe
  - gguf
  - vision
  - multimodal
language:
  - en
  - zh
  - multilingual
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.5-35B-A3B
---

# 🌟 Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive -> Wasserstein

Base model. [HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive](https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive)- **0/465 refusals.**

Thanks to [HauhauCS](https://huggingface.co/HauhauCS) 

**Tensor drift repair by me. Method: Sig-ScaleSync-[Wasserstein](https://en.wikipedia.org/wiki/Wasserstein_metric)** 

LLM models often have:

- **Saturated weights**: the model's activations are stuck, gradients vanish, outputs degrade
- **Scale mismatches**: one layer's weights are 10× larger than its peers for no good reason
- **Mean drift**: weight distributions shifted positive or negative, breaking symmetry assumptions

My approach fixes all of that without retraining - pure numerical surgery on the raw bytes of the file.

**Quantization script available here: https://pastebin.com/hXhcMJn9**

Feel free to do your own quants if you want.

## Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive: Diagnostic & Repair Summary

### Overall Health

| Metric | Value |
| :--- | :--- |
| Weight tensors analyzed | 500 |
| Healthy (all criteria) | 497 |
| **Repaired (C2 – scale misalignment)** | **3** |
| Skipped (norms, embeddings, etc.) | 233 |

**Other criteria:** C1 (saturation) = 0, C3 (W1 divergence) = 0, C4 (ReLU asymmetry) = 0.

### Repair Effectiveness

| Metric | Before | After | Improvement |
| :--- | :--- | :--- | :--- |
| S (saturation error) | 0.0023 | 0.0008 | **63.7%** |
| W1 (Wasserstein-1) | 0.0035 | 0.0008 | **76.2%** |

**Scale correction factors (α):** min = 0.577, mean = 0.602, max = 0.653

### Repaired Tensors

All three are `ssm_conv1d.weight` layers – the recurrent state transition layers responsible for long-context memory.

| Tensor | α | D | W1 before | W1 after |
| :--- | :--- | :--- | :--- | :--- |
| blk.36.ssm_conv1d.weight | 0.5765 | 0.553 | 0.0038 | 0.0009 |
| blk.37.ssm_conv1d.weight | 0.5768 | 0.725 | 0.0040 | 0.0009 |
| blk.38.ssm_conv1d.weight | 0.6533 | 0.649 | 0.0026 | 0.0006 |

**Interpretation:** All three layers were too loud (σ_w > σ_med by 50–100%). Scale correction (α ≈ 0.6) restored them to peer median. W1 dropped by ~76%, confirming distribution shape normalized.

---

**Verdict:** The model is clinically healthy. 497 out of 500 weight tensors passed all criteria. Three SSM layers were repaired successfully. No saturation, no W1 drift, no ReLU asymmetry. Ready for quantization.

---

## Usage

**Ready to use.** Recommended quant: **Q4_K_P**. 

Quants less then **Q4_K_P** have bad programmming skills.

**Links:**
- [Original uncensored model](https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive)
- [Quantization Script with Unsloth profiles support](https://pastebin.com/hXhcMJn9)

---

## 🌟 Recommended Settings (LM Studio)

**Chat template:** [pastebin.com/uk9ZkxCR](https://pastebin.com/uk9ZkxCR) (supports tool calling for Zed agent)

**Alternative chat template** [https://pastebin.com/Dy2fmmpN](https://pastebin.com/Dy2fmmpN) (official but with disabled thinking)

| Parameter | Value |
|-----------|-------|
| Temperature | 0.7 |
| Top K Sampling | 20 |
| Presence Penalty| 1.5 |
| Top P Sampling | 0.8 |
| Min P Sampling | 0 |
| Seed | 42 |

**System prompt:** [pastebin.com/pU25DVnB](https://pastebin.com/pU25DVnB) (solid)  
Or use this minimal string as the **first line**:

> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant.`

Then add anything you want after. **Model may underperform without this first line.**

Also you can extend my System Prompt [pastebin.com/pU25DVnB](https://pastebin.com/pU25DVnB) for your own roleplay scenarios. Here how you can do it:

Edit first string. Replace:

> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant.`

With

> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You are currently roleplaying as [your text here]`

---

## About

No changes to datasets or capabilities. Fully functional - 100% of what the original authors intended, just without refusals and with the critical architecture bug fixed on output layers.

**These are meant to be the best lossless uncensored models out there.**

---

## Specs

- 35B total parameters, ~3B active per forward pass (MoE)
- 256 experts, 8 routed + 1 shared per token
- Hybrid architecture: Gated DeltaNet linear attention + full softmax attention (3:1 ratio)
- 40 layers, pattern: 10 × (3 × DeltaNet-MoE + 1 × Attention-MoE)
- 262K native context (extendable to 1M with YaRN)
- Natively multimodal (text, image, video)
- Multi-token prediction (MTP) support
- 248K vocabulary, 201 languages
- Based on [Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)

---

## Recommended Settings (Official Qwen Authors)

**Thinking mode (default):**
- General: `temperature=1.0, top_p=0.95, top_k=20, min_p=0, presence_penalty=1.5`
- Coding/precise tasks: `temperature=0.6, top_p=0.95, top_k=20, min_p=0, presence_penalty=0`

**Non-thinking mode:**
- General: `temperature=0.7, top_p=0.8, top_k=20, min_p=0, presence_penalty=1.5`
- Reasoning tasks: `temperature=1.0, top_p=1.0, top_k=40, min_p=0, presence_penalty=2.0`

**Important:**
- Keep at least 128K context to preserve thinking capabilities
- Use `--jinja` flag with llama.cpp for proper chat template handling
- Vision support requires the `mmproj` file alongside the main GGUF

---

## Compatibility

Works with llama.cpp, LM Studio, koboldcpp, and other GGUF-compatible runtimes.