ykarout/qwen3.5-9b-opus-openclaw-distilled-gguf overview

{
  "metadata": {},
  "card_data": {
    "tags": [
      "text-generation-inference",
      "transformers",
      "unsloth",
      "qwen3_5",
      "openclaw",
      "agentscope",
      "opus",
      "opus-4.6",
      "qwen",
      "reasoning",
      "chain-of-thought",
      "conversational",
      "unsloth",
      "lora",
      "agentic",
      "tool-use",
      "text-generation"
    ],
    "license": "apache-2.0",
    "language": [
      "en",
      "ar",
      "es",
      "la",
      "tr",
      "de",
      "cs",
      "ch",
      "zh"
    ],
    "datasets": [
      "nohurry/Opus-4.6-Reasoning-3000x-filtered",
      "Roman1111111/claude-opus-4.6-10000x",
      "ykarout/open-r1-sampled",
      "ykarout/Opus-4.6-reasoning-sft-12k"
    ],
    "base_model": [
      "ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled",
      "Qwen/Qwen3.5-9B",
      "agentscope-ai/CoPaw-Flash-9B"
    ],
    "library_name": "transformers",
    "frontmatter": {
      "tags": [
        "text-generation-inference",
        "transformers",
        "unsloth",
        "qwen3_5",
        "openclaw",
        "agentscope",
        "opus",
        "opus-4.6",
        "qwen",
        "reasoning",
        "chain-of-thought",
        "conversational",
        "unsloth",
        "lora",
        "agentic",
        "tool-use",
        "text-generation"
      ],
      "license": "apache-2.0",
      "language": [
        "en",
        "ar",
        "es",
        "la",
        "tr",
        "de",
        "cs",
        "ch",
        "zh"
      ],
      "datasets": [
        "nohurry/Opus-4.6-Reasoning-3000x-filtered",
        "Roman1111111/claude-opus-4.6-10000x",
        "ykarout/open-r1-sampled",
        "ykarout/Opus-4.6-reasoning-sft-12k"
      ],
      "base_model": [
        "ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled",
        "Qwen/Qwen3.5-9B",
        "agentscope-ai/CoPaw-Flash-9B"
      ],
      "library_name": "transformers"
    },
    "hero_image_url": "https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png",
    "summary": "This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.  # Qwen3.5-9B-Opus-OpenClaw-Distilled **Qwen3.5-9B-Opus-OpenClaw-Distilled** is a reasoning-first, agentically-tuned derivative of **Qwen3.5-9b**, built to fuse two strengths into one model identity: The goal is simple: Use a strong base mode --> tune for agentic harness like openclaw and agentscope --> distill opus-4.6 level reasoning --> best of both worlds",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\ntags:\n- text-generation-inference\n- transformers\n- unsloth\n- qwen3_5\n- openclaw\n- agentscope\n- opus\n- opus-4.6\n- qwen\n- reasoning\n- chain-of-thought\n- conversational\n- unsloth\n- lora\n- agentic\n- tool-use\n- text-generation\nlicense: apache-2.0\nlanguage:\n- en\n- ar\n- es\n- la\n- tr\n- de\n- cs\n- ch\n- zh\ndatasets:\n- nohurry/Opus-4.6-Reasoning-3000x-filtered\n- Roman1111111/claude-opus-4.6-10000x\n- ykarout/open-r1-sampled\n- ykarout/Opus-4.6-reasoning-sft-12k\nbase_model:\n- ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled\n- Qwen/Qwen3.5-9B\n- agentscope-ai/CoPaw-Flash-9B\nlibrary_name: transformers\n---\n\n# Uploaded finetuned  model\n\n- **Developed by:** ykarout\n- **License:** apache-2.0\n\n\nThis qwen3_5 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.\n\n[<img src=\"https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png\" width=\"200\"/>](https://github.com/unslothai/unsloth)\n\n\n# Qwen3.5-9B-Opus-OpenClaw-Distilled\n\n**Qwen3.5-9B-Opus-OpenClaw-Distilled** is a reasoning-first, agentically-tuned derivative of **Qwen3.5-9b**, built to fuse two strengths into one model identity:\n\n- **OpenClaw & CoPaw’s operational / agentic instincts**\n- **Claude Opus-style structured reasoning distillation**\n\nThe goal is simple: Use a strong base mode --> tune for agentic harness like openclaw and agentscope --> distill opus-4.6 level reasoning --> best of both worlds \n\n## TL;DR\n\nIt is designed for users who want:\n\n- preserved chat + strong agentic usefulness from the Openclaw / CoPaw lineage\n- a model that feels more “planner + operator” than just “chatbot”\n\nRecommended sampling parameters:\n\n- `temperature=0.6`\n- `top_p=0.95`\n- `min_p=0.0`\n- `top_k=20`\n- `repeat_penalty=1.0`\n- `presence_penalty=0.0`\n\n## Special Instructions for Ollama Only\n\nThe GGUF that works correctly with Ollama is the vision-merged file as Ollama only accepts a single GGUF for loading from Modelfile\n\n- `Qwen3.5-9b-Opus-Openclaw-Distilled-vision-merged-Q6_K.gguf`\n\nThis file already has the multimodal weights merged into the main GGUF, so you should use it directly as the `FROM` target in your `Modelfile`.\n\n### 1. Download the GGUF\n\nIf you want to download it manually from this repo:\n\n```bash\nhuggingface-cli download ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF Qwen3.5-9b-Opus-Openclaw-Distilled-vision-merged-Q6_K.gguf --local-dir .\n```\n\n### 2. Create a `Modelfile`\n\nSave the following as `Modelfile` in the same folder as the GGUF:\n\n```text\nFROM ./Qwen3.5-9b-Opus-Openclaw-Distilled-vision-merged-Q6_K.gguf\nTEMPLATE {{ .Prompt }}\nRENDERER qwen3.5\nPARSER qwen3.5\nPARAMETER temperature 0.6\nPARAMETER top_p 0.95\nPARAMETER min_p 0.0\nPARAMETER top_k 20\nPARAMETER repeat_penalty 1.0\nPARAMETER presence_penalty 0.0\n```\n\n### 3. Create the Ollama model\n\n```bash\nollama create qwen3.5-9b-opus-openclaw:Q6_K -f Modelfile\n```\n\n### 4. Run it\n\n```bash\nollama run qwen3.5-9b-opus-openclaw:Q6_K\n```\n\n### Notes\n\n- Use the `vision-merged` GGUF is for Ollama only. \n- All other GGUFs work out of the box with LMStudio and llama.cpp with text + vision \n\n\n\n\n## Why this model exists\n\n`CoPaw-Flash-9B` is already a highly interesting Qwen3.5-based model family member with explicit optimization for agentic behavior such as tool invocation, command execution, memory management, and multi-step planning. Opus builds on top of that foundation instead of starting from a plain base model. The idea is to preserve that practical “gets things done” behavior while injecting denser and more structured reasoning traces through supervised fine-tuning.\n\nAt the same time, the inspiration for the reasoning side of this model comes from recent Qwen3.5 reasoning distillations trained with Opus-derived trajectories. In particular, models like `Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled` emphasize `<think>`-structured reasoning, response-only training, and normalized reasoning/answer formatting.\n\n## Model identity\n\nThe Model is intended to sit at the intersection of:\n\n- **agentic chat utility**\n- **structured reasoning**\n- **practical local deployment**\n- **Qwen3.5 ecosystem compatibility**\n\nThe intended vibe is:\n\n> **Qwen3.5 + OpenClaw reflexes + Opus-style reasoning scaffolds**\n\nIn practice, that means the model is aimed at tasks like:\n- analytical QA\n- coding support\n- workflow planning\n- terminal / tool-oriented prompting\n- multi-step decomposition\n- logic-heavy conversations\n- “think first, answer second” style interactions\n\n## Base model\n\n- **Base family:** Qwen3.5-9b\n- **Immediate base:** `CoPaw-Flash-9B` --> A fine-tune that exhibits much higher agentic capabilities in harnesses like OpenClaw and AgentScope \n\n## Training concept\n\nThis model was trained as a **text-only reasoning SFT** derivative focused on preserving and reinforcing the format:\n\n```text\n<think>\n...\n</think>\n\nfinal answer\n```\n\nThe overall training philosophy is aligned with reasoning-distillation approaches that emphasize:\n- response-only loss masking\n- explicit `<think>` formatting\n- structured step-by-step reasoning before the final answer\n\n## Datasets\n\nThe training recipe is centered on a unified reasoning dataset built from:\n\n- `Roman1111111/claude-opus-4.6-10000x`\n- `Crownelius/Opus-4.6-Reasoning-3300x`\n\nThese were normalized into a single conversational SFT dataset:\n- `ykarout/Opus-4.6-reasoning-sft-12k`\n\n### Dataset processing highlights\n\nThe dataset was cleaned and unified so both sources follow the same final structure:\n\n- `messages`-based conversational schema\n- assistant output normalized into:\n  - `<think> ... </think>`\n  - followed by the final answer\n- generic repeated system prompts removed where appropriate\n- token-length profile measured after rendering with the target tokenizer chat template\n\nThis makes the training corpus more consistent and more directly usable in TRL / Unsloth conversational SFT pipelines.\n\n## Training recipe\n\nHigh-level recipe:\n\n- **Framework:** Unsloth\n- **Method:** LoRA SFT\n- **Objective:** improve structured reasoning while retaining CoPaw-style usefulness\n- **Loss behavior:** train on assistant responses / completions only\n- **Format target:** explicit `<think>` reasoning followed by answer\n- **Text-only setup:** no vision-layer fine-tuning path used\n\n## What changed versus CoPaw-Flash-9B\n\nThis is not presented as a replacement for CoPaw-Flash-9B’s original design goals.\n\nInstead, it pushes the model further toward:\n- more explicit reasoning traces\n- more deliberate planning language\n- cleaner internal decomposition on complex tasks\n- stronger “reason-then-answer” behavior\n\n\n## Intended use\n\nModel is best suited for:\n- Agentic harnesses like OpenClaw, Claude Code, OpenCode, AgentScope etc..\n- deep analytical prompting\n- code and debugging assistance\n- local agent workflows\n- logic / math / structured breakdown tasks\n\n\nIt is a particularly natural fit for prompts where you want the model to:\n1. parse the task carefully\n2. build a plan\n3. utilize different tools\n4. then produce a clean answer or action\n\n\n\n## Limitations\n\n- This is still an autoregressive language model and can hallucinate.\n- Strong reasoning style does not guarantee factual correctness.\n- More visible reasoning can sometimes increase verbosity.\n- Distillation can improve structure without perfectly reproducing frontier-model judgment.\n- Depending on the prompt mix, some behaviors may lean more “reasoning-first” than “tool-first.”\n\n## Acknowledgements\n\nHuge credit goes to the upstream work that made this possible:\n\n- `agentscope-ai/CoPaw-Flash-9B`\n- `Roman1111111/claude-opus-4.6-10000x`\n- `Crownelius/Opus-4.6-Reasoning-3300x`\n- `Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled`\n- the broader Qwen / Unsloth ecosystem\n\n## Citation / lineage notes\n\nIf you use this model, please also acknowledge the upstream projects and datasets it builds on.\n\n**Qwen3.5-9B-Opus-OpenClaw-Distilled** is for people who want a model that doesn’t just answer — it **locks in, thinks cleanly, and then strikes.**",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "text-generation-inference",
    "unsloth",
    "qwen3_5",
    "openclaw",
    "agentscope",
    "opus",
    "opus-4.6",
    "qwen",
    "reasoning",
    "chain-of-thought",
    "conversational",
    "lora",
    "agentic",
    "tool-use",
    "text-generation",
    "en",
    "ar",
    "es",
    "la",
    "tr",
    "de",
    "cs",
    "ch",
    "zh",
    "dataset:nohurry/Opus-4.6-Reasoning-3000x-filtered",
    "dataset:Roman1111111/claude-opus-4.6-10000x",
    "dataset:ykarout/open-r1-sampled",
    "dataset:ykarout/Opus-4.6-reasoning-sft-12k",
    "base_model:Qwen/Qwen3.5-9B",
    "base_model:adapter:Qwen/Qwen3.5-9B",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us"
  ],
  "likes": 7,
  "downloads": 5524,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-02T10:39:07.000Z",
  "created_at": "2026-04-01T16:56:50.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "transformers"
}

{
  "_id": "69cd4e52f0e4e486defcd4a4",
  "id": "ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF",
  "modelId": "ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF",
  "sha": "1bc0cadcab0a17c18a8fb8f78695414eab4e96fa",
  "createdAt": "2026-04-01T16:56:50.000Z",
  "lastModified": "2026-04-02T10:39:07.000Z",
  "author": "ykarout",
  "downloads": 5524,
  "likes": 7,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "transformers",
  "siblings_count": 9
}