Jackrong/Qwopus3.6-27B-v2
Jackrong/Qwopus3.6-27B-v2 is a 27 billion parameter dense language model, fine-tuned on Qwen3.6-27B by Jackrong, with a 32K context window. It is specifically enhanced for reasoning tasks through a novel Trace Inversion method, which reconstructs detailed step-by-step logical derivations from commercial LLM outputs. This model excels at complex logical reasoning, agentic workflows, and code generation, offering improved reasoning efficiency and token conversion compared to its base model.
Loading preview...
Qwopus3.6-27B-v2: Reasoning-Enhanced Dense Language Model
Qwopus3.6-27B-v2 is a 27 billion parameter dense language model built upon Alibaba Cloud's Qwen3.6-27B, featuring a 32K context window. Developed by Jackrong in collaboration with Kyle Hessling, this model introduces a unique Trace Inversion methodology. This process reverse-engineers compressed "Reasoning Bubbles" from commercial LLMs (like Claude-4.7-Max) into explicit, step-by-step "Learnable Chain-of-Thought (CoT)" traces, effectively eliminating logical shortcuts and knowledge fractures.
Key Capabilities & Features
- Enhanced Reasoning: Achieves 87.43% on a selected MMLU-Pro subset, outperforming Qwen3.6-27B by +2.57 percentage points, particularly strong in Business, Computer Science, Physics, and Chemistry.
- Code Generation & Agentic Workflows: Demonstrates strong performance on SWE-bench with a 75.25% resolution rate and successfully handles complex agentic tasks, including multi-step planning, tool-use, and code debugging.
- Efficiency: Shows significant reasoning efficiency gains, requiring 35.9% fewer tokens on average for correct answers and a 16.6% improvement in token conversion efficiency compared to the base model.
- Vision & Tool-use Support: Natively supports vision and tool-use capabilities, requiring an external
mmproj.gguffor vision functionality. - Curriculum Learning: Trained using a three-stage curriculum to progressively scale reasoning quality under long-context inference, ensuring format stability and robust logical derivations.
Ideal Use Cases
- Complex Problem Solving: Suited for applications requiring deep logical reasoning and structured problem-solving.
- Agentic AI Development: Excellent for building AI agents that need to plan, execute, and self-critique tasks, especially in coding and multi-tool environments.
- Code Generation & Debugging: Strong performance in generating and debugging code, as evidenced by SWE-bench results and creative coding tasks.
- Research & Exploration: Positioned as an experimental community release for research into reasoning distillation and advanced fine-tuning techniques.