Jackrong/Negentropy-claude-opus-4.7-9B

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Jackrong/Negentropy-claude-opus-4.7-9B is a 9 billion parameter reasoning-enhanced experimental model developed by Jackrong, built upon the Qwen3.5-9B base. It utilizes a "Trace Inversion" technique to reconstruct detailed Chains-of-Thought from compressed Claude-Opus-4.7 "Reasoning Bubbles." This model is specifically designed for advanced logical deduction and generating high-quality CoT data, offering superior logical fidelity for complex problem-solving on mid-range hardware.

Loading preview...

Negentropy-claude-opus-4.7-9B: Reasoning-Enhanced Model

Jackrong/Negentropy-claude-opus-4.7-9B is a 9 billion parameter model that addresses the "Information Entropy" trap in model distillation by reconstructing detailed reasoning chains. Developed by Jackrong with support from Kyle Hessling, this model builds on the "Trace Inversion" technique, using the Trace-Inverter-4B to expand compressed "Reasoning Bubbles" from Claude-Opus-4.7 into full Chains-of-Thought (CoT). This process provides high-value supervisory signals for fine-tuning the Qwen3.5-9B base model, allowing for more nuanced logical reconstruction and deeper inheritance of the "Opus style" reasoning depth.

Key Capabilities

  • Superior Logical Fidelity: Offers more stable and detailed thinking chains, particularly in mathematical and coding tasks.
  • Complex Instruction Following: Excels at maintaining multi-faceted constraints throughout long reasoning processes.
  • Native CoT Integration: Seamlessly uses <think>...</think> tags for transparent and auditable reasoning.
  • Optimized Performance: Remains fast and memory-efficient when used with Unsloth and 4-bit quantization.

Good For

  • Advanced Logical Deduction: Solving complex problems requiring sophisticated planning.
  • Reasoning Teacher Model: Generating high-quality CoT data for smaller models (0.8B - 4B).
  • Workstation Inference: Delivering "Pro-level" thinking capabilities on hardware like RTX 3090/4090/5090.
  • Academic Research: Investigating scaling laws of reasoning recovery via trace inversion.

Finetunes
7 models