luffythefox/qwen3.5-35b-a3b-uncensored-wasserstein-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "tags": [
      "uncensored",
      "qwen3.5",
      "moe",
      "gguf",
      "vision",
      "multimodal"
    ],
    "language": [
      "en",
      "zh",
      "multilingual"
    ],
    "pipeline_tag": "image-text-to-text",
    "base_model": "Qwen/Qwen3.5-35B-A3B",
    "frontmatter": {
      "license": "apache-2.0",
      "tags": [
        "uncensored",
        "qwen3.5",
        "moe",
        "gguf",
        "vision",
        "multimodal"
      ],
      "language": [
        "en",
        "zh",
        "multilingual"
      ],
      "pipeline_tag": "image-text-to-text",
      "base_model": "Qwen/Qwen3.5-35B-A3B"
    },
    "hero_image_url": "",
    "summary": "Base model: HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive - **0/465 refusals.** **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. ---",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\ntags:\n  - uncensored\n  - qwen3.5\n  - moe\n  - gguf\n  - vision\n  - multimodal\nlanguage:\n  - en\n  - zh\n  - multilingual\npipeline_tag: image-text-to-text\nbase_model: Qwen/Qwen3.5-35B-A3B\n---\n\n# Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive (Repaired) -> Wasserstein\n\nBase model: [HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive](https://huggingface.co/HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive) - **0/465 refusals.**\n\n**Tensor drift repair by me. Method: Sig-ScaleSync-[Wasserstein](https://en.wikipedia.org/wiki/Wasserstein_metric)** \n\n**Quantization script available here: https://pastebin.com/hXhcMJn9**\n\nFeel free to do your own quants if you want.\n\n---\n\n## Tensor Repair Summary\n\n**Model:** Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-BF16.gguf (64.61 GB)\n\n| Metric | Value |\n|--------|-------|\n| Weight tensors analyzed | 500 |\n| Healthy (all criteria) | 497 |\n| Repaired (C2 - scale misalignment) | 3 |\n| Skipped (norms, embeddings, etc.) | 233 |\n\n**No issues found:** C1 (saturation), C3 (W1 divergence), C4 (ReLU asymmetry)\n\n---\n\n### Repair Statistics\n\n| Metric | Before | After | Improvement |\n|--------|--------|-------|-------------|\n| S (saturation error) | 0.0023 | 0.0009 | 63.0% |\n| W1 (Wasserstein-1) | 0.0034 | 0.0008 | 76.6% |\n\n**Scale repair coefficients (α):** min=0.580, mean=0.608, max=0.657\n\n---\n\n### Repaired Tensors (C2 — scale misalignment)\n\nAll three are `ssm_conv1d.weight` layers - the recurrent state transition layers responsible for long-context memory.\n\n| Block | α | D (log-ratio) | W1 before | W1 after |\n|-------|---|---------------|-----------|----------|\n| blk.36 | 0.5852 | 0.545 | 0.0037 | 0.0009 |\n| blk.37 | 0.5800 | 0.707 | 0.0039 | 0.0009 |\n| blk.38 | 0.6573 | 0.628 | 0.0026 | 0.0006 |\n\n**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.\n\n---\n\n### Peer-Group Statistics (shape: 4×8192)\n\n| Metric | Value |\n|--------|-------|\n| Group size | 30 tensors |\n| Median σ_rob | 0.00652 |\n| Damaged tensors | 3 (blk.36, 37, 38) |\n\n**Scale misalignment distribution:** 3 tensors in [0.50, 1.00) log-ratio range.\n\n---\n\n### Verdict\n\nModel is clinically healthy. 497 out of 500 weight tensors passed all four criteria. Three SSM layers were repaired successfully. No saturation, no KL/W1 drift, no ReLU asymmetry. Ready for quantization.\n\n---\n\n## Usage\n\n**Ready to use.** Recommended quantization: **Q4_K_L, or higher** (Q4_K_M, Q5_K_M, Q6_K, Q8_0).  \n⚠️ Lower formats (Q3_K, Q2_K) break the model due to MoE + DeltaNet sensitivity.\n\n**Links:**\n- [Original uncensored model](https://huggingface.co/HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive)\n- [Quantization Script with Unsloth profiles support](https://pastebin.com/hXhcMJn9)\n\n---\n\n## 🌟 Recommended Settings (LM Studio)\n\n**Chat template:** [pastebin.com/uk9ZkxCR](https://pastebin.com/uk9ZkxCR) (supports tool calling for Zed agent)\n\n| Parameter | Value |\n|-----------|-------|\n| Temperature | 0.7 |\n| Top K Sampling | 20 |\n| Presence Penalty | 1.5 |\n| Top P Sampling | 0.8 |\n| Min P Sampling | 0 |\n| Seed | 42 |\n\n**System prompt:** [pastebin.com/pU25DVnB](https://pastebin.com/pU25DVnB) (solid)  \nOr use this minimal string as the **first line**:\n\n> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant.`\n\nThen add anything you want after. **Model may underperform without this first line.**\n\nAlso you can extend my System Prompt [pastebin.com/pU25DVnB](https://pastebin.com/pU25DVnB) for your own roleplay scenarios. Here how you can do it:\n\nEdit first string. Replace:\n\n> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant.`\n\nWith\n\n> `You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You are currently roleplaying as [your text here]`\n\n---\n\n## About\n\nNo 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.\n\n**These are meant to be the best lossless uncensored models out there.**\n\n---\n\n## Specs\n\n- 35B total parameters, ~3B active per forward pass (MoE)\n- 256 experts, 8 routed + 1 shared per token\n- Hybrid architecture: Gated DeltaNet linear attention + full softmax attention (3:1 ratio)\n- 40 layers, pattern: 10 × (3 × DeltaNet-MoE + 1 × Attention-MoE)\n- 262K native context (extendable to 1M with YaRN)\n- Natively multimodal (text, image, video)\n- Multi-token prediction (MTP) support\n- 248K vocabulary, 201 languages\n- Based on [Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)\n\n---\n\n## Recommended Settings (Official Qwen Authors)\n\n**Thinking mode (default):**\n- General: `temperature=1.0, top_p=0.95, top_k=20, min_p=0, presence_penalty=1.5`\n- Coding/precise tasks: `temperature=0.6, top_p=0.95, top_k=20, min_p=0, presence_penalty=0`\n\n**Non-thinking mode:**\n- General: `temperature=0.7, top_p=0.8, top_k=20, min_p=0, presence_penalty=1.5`\n- Reasoning tasks: `temperature=1.0, top_p=1.0, top_k=40, min_p=0, presence_penalty=2.0`\n\n**Important:**\n- Keep at least 128K context to preserve thinking capabilities\n- Use `--jinja` flag with llama.cpp for proper chat template handling\n- Vision support requires the `mmproj` file alongside the main GGUF\n\n---\n\n## Compatibility\n\nWorks with llama.cpp, LM Studio, koboldcpp, and other GGUF-compatible runtimes.",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "uncensored",
    "qwen3.5",
    "moe",
    "vision",
    "multimodal",
    "image-text-to-text",
    "en",
    "zh",
    "multilingual",
    "base_model:Qwen/Qwen3.5-35B-A3B",
    "base_model:quantized:Qwen/Qwen3.5-35B-A3B",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 1,
  "downloads": 2188,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-16T15:14:15.000Z",
  "created_at": "2026-04-16T08:47:53.000Z",
  "pipeline_tag": "image-text-to-text",
  "library_name": ""
}

{
  "_id": "69e0a23952d270b64fb2bb9d",
  "id": "LuffyTheFox/Qwen3.5-35B-A3B-Uncensored-Wasserstein-GGUF",
  "modelId": "LuffyTheFox/Qwen3.5-35B-A3B-Uncensored-Wasserstein-GGUF",
  "sha": "8a5cb729b040da007534237ac263b22b19e28b15",
  "createdAt": "2026-04-16T08:47:53.000Z",
  "lastModified": "2026-04-16T15:14:15.000Z",
  "author": "LuffyTheFox",
  "downloads": 2188,
  "likes": 1,
  "gated": false,
  "private": false,
  "pipeline_tag": "image-text-to-text",
  "library_name": "",
  "siblings_count": 5
}