license: apache-2.0
tags:
- uncensored
- qwen3.6
- moe
- gguf
- vision
- multimodal
- genesis
- hermes
- agentic
language: - en
- zh
- multilingual
datasets: - NousResearch/hermes-function-calling-v1
pipeline_tag: image-text-to-text
base_model: - HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
⚡ https://web.tribute.tg/d/KIH ⚡ If you like this Genesis LLM release you can donate to me via @Tribute bot in Telegram messenger and support future Genesis LLM development.
🌟 Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive -> Genesis Hermes V2
Key diffrence from previous V1 release is data reconstruction with noise supression in ssm_out.weight, attn_output.weight and attn_gate.weight tensors via SVD.
Mine approach based on data reconstruction in model via mathematical statistics. I scan blocks in model via chunks via 3 parameters and pick best one that fits to weight distribution in tensor. Best picked chunk replaces zero chunks in broken tensor without touching learned structure. Scanning works on tensors with same name and shape. ssm_conv1d tensors are fixed via alpha multiply for full tensor. I scan all ssm_conv1d tensors weight and scale distribution and normalize scale for weights only for too loud tensors.
Model is based on HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive base.
And DJLougen/hermes-qwen3.5-35b-a3b-GGUF finetune for Hermes agent.
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
Usage
Ready to use. Recommended quant: APEX or Q8_K_P
Tensor drift repair by me. Method: Sig-ScaleSync-Genesis-SVD
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive: Diagnostic & Repair Summary
| Metric | Value |
|---|---|
| Weight tensors analyzed | 500 |
| Healthy (all criteria) | 497 |
| Repaired (C2 – scale misalignment) | 3 |
| Skipped | 233 |
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 – recurrent state transition layers responsible for long‑context memory.
| 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 restored them to peer median. W1 dropped by ≈80%, confirming distribution shape normalized.
Verdict: Model is clinically healthy. 497 out of 500 weight tensors passed all four criteria. Three SSM layers repaired successfully. No saturation, no W1 drift, no ReLU asymmetry. Ready for use.
Links:
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
- Zero blocks: zero blocks corrupt the signal, turning training into noise amplification.
- Training Noise: training noise increase randomness and ruins model output quality.
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.
Any questions?
Contact: [email protected]
My Telegram: @LuffyTheFox
Recommended Settings for RTX 3060 12 GB for best perfomance on APEX quant
Chat template: chat_template.jinja
Set K Cache Quantization Type and V Cache Quantization Type to Q8_0.
Set Number of layers for which to force MoE weights onto CPU to 25.
Set GPU offload to 15. Set seed to 42. Set number of active experts to 9.
For best stability I recommend starting from this simple string in System Prompt with enabled thinking and nothing else:
You are Qwen, a large language model created by Tongyi Lab team from Alibaba Group. You are a helpful assistant.
Thinking mode (default):
- Coding/precise tasks:
temperature=0.6, top_p=0.95, top_k=20, min_p=0, presence_penalty=disabled, repeat_penalty=disabled - General:
temperature=1.0, top_p=0.95, top_k=20, min_p=0, presence_penalty=disabled
For roleplay i recommend to disable thinking and use your own system prompt with following settings:
Non-thinking mode:
- General:
temperature=0.7, top_p=0.8, top_k=20, min_p=0, presence_penalty=disabled, repeat_penalty=disabled
After configuration use this prompt for benchmark: Generate an SVG of a pelican riding a bicycle
I am using Coding/precise tasks preset with Thinking mode enabled
I got this result: pelican.svg
Important:
- Keep at least 128K context to preserve thinking capabilities
- Use
--jinjaflag with llama.cpp for proper chat template handling - Vision support requires the
mmprojfile alongside the main GGUF
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)
- 248K vocabulary, 201 languages
- Base model. HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
Compatibility
Works with llama.cpp, LM Studio, koboldcpp, and other GGUF-compatible runtimes.