Qwen3.6-35B-A3B-Uncensored-Wasserstein
Tensor drift repair by me. Method: Sig-ScaleSync-Wasserstein
Quantization script available here: https://pastebin.com/hXhcMJn9
Feel free to do your own quants if you want.
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 – located in deep blocks (36–38).
| Tensor | α | D (log-ratio) | 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.
Usage
Ready to use. Recommended quantization: Q4_K_L, or higher (Q4_K_M, Q5_K_M, Q6_K, Q8_0).
⚠️ Lower formats (Q3_K, Q2_K) break the model due to MoE + DeltaNet sensitivity.
Links:
🌟 Recommended Settings (LM Studio)
Chat template: pastebin.com/uk9ZkxCR (supports tool calling for Zed agent)
| 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 (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 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
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.
