snowman0919/Qwopus3.6-27B-v2-heretic

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 4, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

Qwopus3.6-27B-v2-heretic is a 27 billion parameter, decensored version of the Jackrong/Qwopus3.6-27B-v2 model, built on the Qwen3.6-27B architecture with a 32K context length. It is fine-tuned using 'Trace Inversion' to reconstruct detailed reasoning pathways from commercial LLM outputs, significantly enhancing logical reasoning, coding, and agentic capabilities. This model excels at complex problem-solving by providing explicit, step-by-step derivations, making it suitable for applications requiring robust and transparent reasoning.

Loading preview...

What the fuck is this model about?

Qwopus3.6-27B-v2-heretic is a 27 billion parameter language model, a decensored variant of Jackrong/Qwopus3.6-27B-v2. It's built upon Alibaba Cloud's Qwen3.6-27B base model, which natively supports long-context modeling up to 32K/128K tokens and is designed for agentic workflows, complex logical reasoning, and multimodal tasks (vision & tool-use).

What makes THIS different from all the other models?

This model's primary differentiator is its Trace Inversion training methodology. Unlike traditional distillation that struggles with compressed "Reasoning Bubbles" from commercial LLMs, Qwopus3.6-27B-v2-heretic uses a specialized logical reconstructor (Trace-Inverter-4B) to reverse-engineer these bubbles into complete, step-by-step "Learnable Chain-of-Thought (CoT)" traces. This process eliminates logical shortcuts and knowledge fractures, leading to:

  • Enhanced Reasoning Depth: Achieves 87.43% on a selected MMLU-Pro subset, outperforming the original Qwen3.6-27B (84.86%), with notable gains in Physics (+10 pp) and Chemistry (+6 pp).
  • Improved Code Generation: Achieved 75.25% on SWE-bench (controlled-202 slice), resolving 152 out of 202 issues.
  • Efficiency in Reasoning: Demonstrates a 35.9% reduction in average output tokens for correctly answered questions compared to the base model, indicating more concise and direct reasoning.
  • Vision & Tool-use Support: Natively supports vision and tool-use capabilities, requiring an mmproj.gguf file for vision functionality.

Should I use this for my use case?

Use this model if your application requires:

  • Complex Logical Reasoning: Ideal for tasks demanding explicit, step-by-step derivations, such as mathematical problem-solving, scientific inquiry, or intricate logical puzzles.
  • Robust Code Generation & Debugging: Its strong performance on SWE-bench and ability to handle creative coding tasks make it suitable for developer tools, automated code review, or agentic coding workflows.
  • Agentic Workflows: Optimized for long-context reasoning and tool-use, making it a strong candidate for building sophisticated AI agents.
  • Reduced Refusals: With 6/100 refusals compared to the original model's 88/100, it's designed for use cases where a less restrictive output is desired.

Consider alternatives if:

  • You need maximum raw throughput for simple generations, where a MoE model might offer higher tokens/second (e.g., Qwopus 3.6 35B-A3B MoE).
  • Your use case is highly sensitive to experimental community releases, as this model has not undergone complete safety evaluations or standard benchmarking.

Finetunes
2 models