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

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled is a 9 billion parameter Qwen3.5-based model developed by ykarout, fine-tuned for enhanced reasoning and agentic capabilities. It integrates the operational instincts of OpenClaw/CoPaw with Claude Opus-style structured reasoning distillation. This model excels at analytical QA, coding support, workflow planning, and multi-step decomposition, making it suitable for agentic workflows and logic-heavy conversations.

Loading preview...

Model Overview

ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled is a 9 billion parameter model derived from Qwen3.5-9b, developed by ykarout. It uniquely combines the agentic strengths of the OpenClaw and CoPaw lineages with advanced, Claude Opus-style structured reasoning. The model was trained using Unsloth and Huggingface's TRL library, focusing on a "think first, answer second" interaction style.

Key Capabilities

  • Agentic Utility: Preserves and enhances agentic usefulness from the OpenClaw/CoPaw lineage, including tool invocation, command execution, memory management, and multi-step planning.
  • Structured Reasoning: Incorporates Opus-style reasoning distillation, emphasizing explicit <think> formatting and structured step-by-step reasoning before a final answer.
  • Practical Local Deployment: Designed for efficient local deployment, offering a balance of performance and accessibility.
  • Qwen3.5 Ecosystem Compatibility: Built upon the Qwen3.5 base, ensuring compatibility within its broader ecosystem.

Intended Use Cases

This model is particularly well-suited for:

  • Agentic harnesses like OpenClaw, Claude Code, OpenCode, and AgentScope.
  • Deep analytical prompting and complex problem-solving.
  • Coding and debugging assistance.
  • Local agent workflows requiring multi-step decomposition.
  • Logic-heavy conversations and tasks demanding structured breakdown.

Training Details

The model was trained as a text-only reasoning SFT derivative, utilizing a unified reasoning dataset from Roman1111111/claude-opus-4.6-10000x and Crownelius/Opus-4.6-Reasoning-3300x. The training focused on response-only loss masking and explicit <think> formatting to reinforce structured reasoning.