Jackrong/Qwen3.5-9B-Neo

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jackrong/Qwen3.5-9B-Neo is a 9 billion parameter, reasoning-focused fine-tune of the Qwen3.5 model, designed to enhance structured multi-step reasoning and mathematical performance. It features an optimized reasoning scaffold that improves logical deduction and problem-solving efficiency. This model excels in tasks requiring complex thought processes, such as competitive programming and advanced mathematics, while maintaining a 32768 token context length.

Loading preview...

Jackrong/Qwen3.5-9B-Neo: Enhanced Reasoning Model

Jackrong/Qwen3.5-9B-Neo is a 9 billion parameter model, fine-tuned from Qwen3.5, with a primary focus on significantly improving reasoning and mathematical capabilities. This "Neo" iteration introduces a highly optimized reasoning scaffold that aims to "think smarter, not longer" by eliminating redundant internal loops and circular reasoning.

Key Capabilities & Differentiators

  • Optimized Reasoning: Features a unique reasoning scaffold designed for structured thought processes, anchoring intermediate steps and improving multi-hop logical consistency.
  • Performance Gains: Achieves notable improvements on critical reasoning benchmarks, including +0.87 pp on BBH, +0.98 pp on MATH Hard, and a significant +2.91 pp on MUSR (multi-step reasoning under uncertainty), as evaluated on the LM Evaluation Harness leaderboard.
  • Structured Output: Conditioned to explicitly structure its thought processes within <think>...</think> tags, providing transparent and methodical problem-solving.
  • Specialized Training Data: Trained on high-quality, filtered reasoning distillation data, including stepfun-ai/Step-3.5-Flash-SFT and a custom Competitive-Programming-python-blend dataset.

Intended Use Cases

  • Offline Analytical Tasks: Ideal for scenarios requiring deep analysis and logical deduction.
  • Coding & Competitive Programming: Excels in generating structured solutions for complex programming challenges.
  • Mathematics: Strong performance in hard mathematical problems.
  • Heavy Logic-Dependent Prompting: Useful when transparent, step-by-step internal logic is crucial for the user.

Finetunes
3 models