Sign up

khazarai/Qwen3.5-2B-Qwen3.6-plus-Distilled-GGUF

🤗 Hugging Face sourceimage-text-to-textapache-2.05.8 GBGGUF✓ HF checksums availableupdated today
No torrent yet →

Model Card: Qwen3.5-Qwen3.6-plus-Reasoning-Distilled-GGUF

Overview

This model is a distilled reasoning-enhanced variant of Qwen3.5-2B, designed to improve:

  • Structured reasoning
  • Step-by-step problem solving
  • Decision stability
  • Output efficiency (token usage)

The model is trained via distillation from a stronger reasoning model (Qwen3.6-plus), transferring:

  • Clean reasoning trajectories
  • Better stopping behavior
  • Reduced reasoning noise

Key Improvements Over Base Model

Reasoning Efficiency

Compared to the base model, this model:

  • Produces shorter and more relevant reasoning chains
  • Avoids repetitive self-verification loops
  • Maintains high signal-to-noise ratio

Stability

The base model often exhibits:

  • Overthinking
  • Infinite or near-infinite reasoning loops
  • Hypothesis explosion

This distilled model:

  • Converges faster to a solution
  • Maintains deterministic reasoning paths
  • Avoids reasoning drift

Decision-Making

  • Improved reasoning termination policy
  • Clearer final answers
  • Better alignment between reasoning and output

Known Failure Modes

  • Occasional hallucinated justifications
  • Overconfidence in incorrect options
  • Missing rare edge-case interpretations
  • Limited deep domain reasoning beyond training distribution

Available Model files:

qwen3.6-plus-Distilled-GGUF.F16.gguf qwen3.6-plus-Distilled-GGUF.Q8.gguf

An Ollama Modelfile is included for easy deployment.