YOYO-AI/Qwen3.5-9B-YOYO-Thinking

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3.5-9B-YOYO-Thinking is a 9 billion parameter language model developed by YOYO-AI, based on the Qwen3.5 architecture. It utilizes an Optimal Transport Merge method, derived from recent research, to combine different model sizes. This model is notable for its significantly extended context length of 262,144 tokens and is designed for general and coding tasks, operating with bfloat16 precision.

Loading preview...

Model Overview

YOYO-AI/Qwen3.5-9B-YOYO-Thinking is a 9 billion parameter model built upon the Qwen3.5 architecture. Its development incorporates a novel cross-model size merging technique, specifically the Optimal Transport Merge method, which is based on recent research papers. This approach aims to enhance model capabilities by effectively combining insights from different model scales.

Key Features and Capabilities

  • Architecture: Based on the Qwen3.5 family.
  • Merging Method: Employs Optimal Transport Merge for cross-model size integration.
  • Precision: Operates with bfloat16 data type.
  • Extended Context Length: Features an impressive context window of 262,144 tokens, significantly larger than many comparable models.
  • Task Optimization: Provides distinct parameter settings for general tasks and coding tasks, suggesting specialized tuning for different applications.
    • General Tasks: Recommended settings include temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0.
    • Coding Tasks: Recommended settings include temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0.

Training Details

The model's development involved activation extraction using 2,000 prompts from the CodeAlpaca_20K dataset, indicating a focus on code-related understanding and generation. Further technical details regarding the merging process and tensor states are available in merge_stats.csv and global_state.json within the repository.

Research and Development

This model is a direct implementation of methodologies described in a research paper and its development is detailed in the associated GitHub Repository.

Finetunes
1 models