CoolFace
Modelpublic

ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
4likes41downloads
Model Card

Uploaded finetuned model

  • —Developed by: ykarout
  • —License: apache-2.0

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

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:

  • —OpenClaw & CoPaw’s operational / agentic instincts
  • —Claude Opus-style structured reasoning distillation

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

GGUFs at ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF

TL;DR

It is designed for users who want:

  • —preserved chat + strong agentic usefulness from the Openclaw / CoPaw lineage
  • —a model that feels more “planner + operator” than just “chatbot”

Why this model exists

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.

At 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.

Model identity

The Model is intended to sit at the intersection of:

  • —agentic chat utility
  • —structured reasoning
  • —practical local deployment
  • —Qwen3.5 ecosystem compatibility

The intended vibe is:

Qwen3.5 + OpenClaw reflexes + Opus-style reasoning scaffolds

In practice, that means the model is aimed at tasks like:

  • —analytical QA
  • —coding support
  • —workflow planning
  • —terminal / tool-oriented prompting
  • —multi-step decomposition
  • —logic-heavy conversations
  • —“think first, answer second” style interactions

Base model

  • —Base family: Qwen3.5-9b
  • —Immediate base: CoPaw-Flash-9B --> A fine-tune that exhibits much higher agentic capabilities in harnesses like OpenClaw and AgentScope

Training concept

This model was trained as a text-only reasoning SFT derivative focused on preserving and reinforcing the format:

text
<think>
...
</think>

final answer

The overall training philosophy is aligned with reasoning-distillation approaches that emphasize:

  • —response-only loss masking
  • —explicit <think> formatting
  • —structured step-by-step reasoning before the final answer

Datasets

The training recipe is centered on a unified reasoning dataset built from:

  • —Roman1111111/claude-opus-4.6-10000x
  • —Crownelius/Opus-4.6-Reasoning-3300x

These were normalized into a single conversational SFT dataset:

  • —ykarout/Opus-4.6-reasoning-sft-12k

Dataset processing highlights

The dataset was cleaned and unified so both sources follow the same final structure:

  • —messages-based conversational schema
  • —assistant output normalized into:
  • —<think> ... </think>
  • —followed by the final answer
  • —generic repeated system prompts removed where appropriate
  • —token-length profile measured after rendering with the target tokenizer chat template

This makes the training corpus more consistent and more directly usable in TRL / Unsloth conversational SFT pipelines.

Training recipe

High-level recipe:

  • —Framework: Unsloth
  • —Method: LoRA SFT
  • —Objective: improve structured reasoning while retaining CoPaw-style usefulness
  • —Loss behavior: train on assistant responses / completions only
  • —Format target: explicit <think> reasoning followed by answer
  • —Text-only setup: no vision-layer fine-tuning path used

What changed versus CoPaw-Flash-9B

This is not presented as a replacement for CoPaw-Flash-9B’s original design goals.

Instead, it pushes the model further toward:

  • —more explicit reasoning traces
  • —more deliberate planning language
  • —cleaner internal decomposition on complex tasks
  • —stronger “reason-then-answer” behavior

Intended use

Model is best suited for:

  • —Agentic harnesses like OpenClaw, Claude Code, OpenCode, AgentScope etc..
  • —deep analytical prompting
  • —code and debugging assistance
  • —local agent workflows
  • —logic / math / structured breakdown tasks

It is a particularly natural fit for prompts where you want the model to:

  1. 1.parse the task carefully
  2. 2.build a plan
  3. 3.utilize different tools
  4. 4.then produce a clean answer or action

Limitations

  • —This is still an autoregressive language model and can hallucinate.
  • —Strong reasoning style does not guarantee factual correctness.
  • —More visible reasoning can sometimes increase verbosity.
  • —Distillation can improve structure without perfectly reproducing frontier-model judgment.
  • —Depending on the prompt mix, some behaviors may lean more “reasoning-first” than “tool-first.”

Acknowledgements

Huge credit goes to the upstream work that made this possible:

  • —agentscope-ai/CoPaw-Flash-9B
  • —Roman1111111/claude-opus-4.6-10000x
  • —Crownelius/Opus-4.6-Reasoning-3300x
  • —Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
  • —the broader Qwen / Unsloth ecosystem

Citation / lineage notes

If you use this model, please also acknowledge the upstream projects and datasets it builds on.

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.