Jackrong/Qwopus3.6-27B-Coder

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Jackrong/Qwopus3.6-27B-Coder is a 27 billion parameter dense transformer model built on Qwen3.6-27B, specialized for agentic coding and tool-use reasoning. Fine-tuned using a multi-stage curriculum and Trace Inversion, it excels at repository-level coding tasks, multi-turn tool orchestration, and complex logical reasoning. The model achieves 67.0% on SWE-bench Verified in a thinking-off mode, optimized for fast local agentic coding workflows with a 32K token context window.

Loading preview...

What is Qwopus-3.6-27B-Coder?

Qwopus-3.6-27B-Coder is a 27 billion parameter dense language model developed by Jackrong, built upon the Qwen3.6-27B base model from Alibaba Cloud. It is specifically fine-tuned for agentic coding, structured tool calling, debugging, and instruction-following within developer workflows. The model inherits a strong reasoning foundation from its base, which achieved 87.43% MMLU-Pro and 75.25% SWE-bench Verified.

Key Capabilities & Differentiators

  • Agentic Coding: Optimized for repository-level coding, debugging, patch generation, and multi-step development workflows.
  • Tool Calling: Learns from real agent trajectories, including tool definitions, calls, and environment feedback for robust multi-turn execution.
  • Trace Inversion: Utilizes a unique distillation strategy to reconstruct compressed "Reasoning Bubbles" from models like Claude Opus into full, learnable step-by-step reasoning traces, addressing the "Information Entropy Trap."
  • 27B Scale & Deployability: Offers deep reasoning with 27 billion parameters and native 32K context support, while remaining practical for single-GPU deployment (e.g., RTX 5090).
  • High Performance (No-Thinking Mode): Achieved 67.0% on the full SWE-bench Verified benchmark (335/500 tasks resolved) in a "thinking-off" mode, demonstrating strong practical coding ability at approximately 100 tokens/sec on an RTX 5090 with MTP enabled.

When to Use This Model

  • Agentic Code Generation: For tasks requiring the model to interact with environments, debug code, or generate patches.
  • Tool Orchestration: When complex, multi-turn tool use and structured instruction following are critical.
  • Local Development Workflows: Ideal for developers seeking a powerful, locally deployable coding agent that prioritizes speed and direct action quality over explicit, long reasoning traces.
  • Repository-Level Tasks: Excels at understanding and modifying code within larger project structures.

Finetunes
6 models