Jackrong/Qwen3.5-9B-Gemini-3.1-Pro-Reasoning-Distill

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jackrong/Qwen3.5-9B-Gemini-3.1-Pro-Reasoning-Distill is a 9 billion parameter reasoning model built upon the Qwen3.5-9B architecture, fine-tuned with high-density reasoning distillation from Gemini 3.1, Qwen3.5-27B, and Gemini 3.0 Pro. It is specifically optimized for structured analytical behavior, enhancing decomposition, planning, abstraction, and response cleanliness in complex multi-step tasks. This model excels at providing coherent, organized, and high-density Chain-of-Thought (CoT) reasoning across diverse domains.

Loading preview...

Model Overview

Jackrong/Qwen3.5-9B-Gemini-3.1-Pro-Reasoning-Distill is a 9 billion parameter language model, fine-tuned from the Qwen3.5-9B base model. Its core innovation lies in its reasoning distillation pipeline, primarily leveraging high-density reasoning traces from Gemini 3.1, supplemented by data from Qwen3.5-27B and Gemini 3.0 Pro. This Supervised Fine-Tuning (SFT) process aims to instill a more structured, coherent, and high-density Chain-of-Thought (CoT) reasoning pattern.

Key Capabilities

  • Structured Analytical Reasoning: Optimized to identify and organize problem structures before generating responses, moving beyond shallow completion.
  • Improved Multi-Step Planning: Demonstrates enhanced performance on tasks requiring decomposition, constraint tracking, sequential planning, and trade-off analysis.
  • Cross-Domain Reasoning: Trained on a diverse corpus covering mathematics, programming, systems, physics, law, medicine, finance, and chemistry, ensuring broad analytical consistency.
  • Security & Adversarial Awareness: Includes distilled data on adversarial and failure-mode reasoning, improving robustness against difficult prompts.
  • Compact but Powerful: Delivers significantly denser reasoning behavior and cleaner analytical output compared to generic instruct models of similar size.

Training & Distillation

The model's training involved SFT with LoRA and reasoning distillation, specifically masking for response-only training on assistant turns. Datasets included Roman1111111/gemini-3.1-pro-hard-high-reasoning for structured analytical style, Jackrong/Qwen3.5-reasoning-700x for Qwen-family reasoning trajectories, and Roman1111111/gemini-3-pro-10000x-hard-high-reasoning for broad multi-domain coverage. This process aims to transfer cleaner analytical structure and stronger planning habits.

Limitations

Users should be aware of potential hallucination risks for niche facts, a reasoning style bias towards longer, more structured answers even for simple prompts, and a teacher-style distillation bias reflecting the source models' reasoning patterns. As a specialized distilled model, prompt formatting and ecosystem integrations may require specific tuning.