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

Genesis - 7.5M dead blocks across 55 tensors were restored from each tensor's own live-weight distribution. Signal flow between layers has been calibrated. Early expert layers (blocks 0–1) were the most affected and are now fully recovered. The model shows healthy tensor statistics, intact expert diversity, and no signs of degradation from the restoration process. ssm1_conv1d drift in layers also has been fixed in this release. Now model can be finetuned.

Here diffrence between Wasserstein and Genesis versions: link

Usage

Recommended quant: MXFP4_MOE.

Join the Discord for updates, roadmaps, projects, or just to chat.

Base model. HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive- 0/465 refusals.

Thanks to HauhauCS

Tensor drift and zero blocks repair by me. Method: Sig-ScaleSync-Genesis

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.

Model Restoration Report

Metric	Value
Total blocks restored	7,528,285
Tensors recovered	55
Scale corrections	3
Expert collapse	0
Reverts (worse after repair)	0
Distribution source	100% own-distribution

---

Restored tensors

All recovered tensors used their own live-weight distribution — no external averaging.

#	Tensor	Blocks restored
1	blk.0.ffn_gate_exps	3,472,448
2	blk.0.ffn_up_exps	3,472,448
3	blk.1.attn_qkv	13,184
4	blk.1.ffn_gate_exps	137,792
5	blk.1.ffn_up_exps	137,792
6	blk.2.attn_qkv	7,296
7	blk.2.ffn_gate_exps	60,928
8	blk.2.ffn_up_exps	60,928
9	blk.3.ffn_gate_exps	23,680
10	blk.3.ffn_up_exps	23,680
11	blk.4.ffn_gate_exps	11,584
12	blk.4.ffn_up_exps	11,584
13	blk.5.ffn_gate_exps	1,152
14	blk.5.ffn_up_exps	1,152
15	blk.6.ffn_gate_exps	128
16	blk.6.ffn_up_exps	128
17	blk.7.ffn_gate_exps	128
18	blk.7.ffn_up_exps	128
19	blk.8.ffn_gate_exps	1,536
20	blk.8.ffn_up_exps	1,536
21	blk.8.ssm_alpha	47
22	blk.9.attn_qkv	128
23	blk.9.ffn_gate_exps	1,856
24	blk.9.ffn_up_exps	1,856
25	blk.10.attn_qkv	256
26	blk.10.ffn_gate_exps	64
27	blk.10.ffn_up_exps	64
28	blk.11.ffn_gate_exps	13,568
29	blk.11.ffn_up_exps	13,568
30	blk.13.ffn_gate_exps	2,851
31	blk.13.ffn_up_exps	2,827
32	blk.14.ffn_gate_exps	1,920
33	blk.14.ffn_up_exps	1,920
34	blk.15.ffn_gate_exps	576
35	blk.15.ffn_up_exps	576
36	blk.16.ffn_gate_exps	448
37	blk.16.ffn_up_exps	448
38	blk.17.ffn_gate_exps	2,240
39	blk.17.ffn_up_exps	2,240
40	blk.18.ffn_gate_exps	576
41	blk.18.ffn_up_exps	576
42	blk.21.attn_qkv	128
43	blk.21.ffn_gate_exps	1,856
44	blk.21.ffn_up_exps	1,856
45	blk.22.attn_qkv	256
46	blk.22.ffn_gate_exps	64
47	blk.22.ffn_up_exps	64
48	blk.23.ffn_gate_exps	13,568
49	blk.23.ffn_up_exps	13,568
50	blk.25.ffn_gate_exps	2,176
51	blk.25.ffn_up_exps	2,176
52	blk.26.ffn_gate_exps	1,920
53	blk.26.ffn_up_exps	1,920
54	blk.28.ffn_gate_exps	448
55	blk.28.ffn_up_exps	448

Scale-adjusted tensors

Three SSM convolution layers received minor scale realignment. All passed verification.

Tensor	Correction
blk.36.ssm_conv1d	α = 0.5765
blk.37.ssm_conv1d	α = 0.5765
blk.38.ssm_conv1d	α = 0.6500

Quality indicators

• Saturation error reduced by 63.8%
• Distributional distance reduced by 76.6%
• All 58 modified tensors passed verification
• No expert representation collapse detected

---

Automated statistical recovery. No retraining, no weight merging, no external data.

Links:

• Original uncensored model
• Quantization Script with Unsloth profiles support

---

Wanna fix your GGUF model?

Contact: luffythefox@mail.ru

My Telegram: @LuffyTheFox

🌟 Recommended Settings (LM Studio)

Chat template: chat_template.jinja

If you want to enable thinking process set enable_thinking variable and preserve_thinking from False to True in chat template. For complex tasks model works a lot better with thinking enabled.

Thanks to froggeric

Parameter	Value
Temperature	0.7
Top K Sampling	20
Presence Penalty	1.5
Repeat Penaly	Disabled
Top P Sampling	0.8
Min P Sampling	0
Seed	42

System prompt: System_Prompt.txt

Benchmark prompt: benchmark.txt

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.txt 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
• Base model. HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

---

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.