horiuchinobuyuki/Qwick-3.5-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwick-3.5-9B is a 9 billion parameter Qwen3.5 derivative developed by Nobuyuki Horiuchi, fine-tuned to produce shorter reasoning traces while maintaining answer quality. This model excels in text reasoning, instruction following, Japanese question answering, and code generation, offering significantly reduced completion lengths across various benchmarks. It supports a 32,768-token context length and is primarily intended for research and evaluation.

Loading preview...

Overview

Qwick-3.5-9B is a 9 billion parameter model developed by Nobuyuki Horiuchi, derived from Qwen3.5-9B. Its primary differentiation lies in its optimization for shorter reasoning traces, aiming to reduce the number of "thinking tokens" without compromising answer quality. This fine-tuned model achieves a noticeable reduction in mean completion length (20% to 55%) across various benchmarks compared to its base model, Qwen3.5-9B, while largely preserving or even improving accuracy.

Key Capabilities and Performance

  • Efficient Reasoning: Significantly reduces completion length for reasoning tasks, making it more efficient.
  • Strong Benchmark Performance: Maintains high accuracy on MMLU-Pro, GPQA-Diamond, and IFEval, with notable improvements in LiveCodeBench (+7.299 pp).
  • Multimodal Retention: Demonstrates strong retention of vision capabilities on the MMMU validation split, showing no degradation in BF16 vision measurement.
  • Code Generation: Shows improved performance in code generation tasks, as evidenced by LiveCodeBench results.
  • Multilingual Support: Includes capabilities for Japanese question answering.

Intended Use Cases

  • Research and Evaluation: Ideal for studying text reasoning, instruction following, and code generation.
  • Applications Requiring Concise Responses: Suitable for scenarios where shorter, more direct answers are preferred without sacrificing accuracy.
  • Japanese Language Processing: Can be utilized for tasks involving Japanese question answering.
  • Code Development: Useful for code generation and related programming tasks.

Finetunes
5 models