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

Based on the latest research papers, we have successfully implemented cross-model size merging and applied the paper's methodology to the latest Qwen3.5 architecture.

Model Highlights:

• merge method: Optimal Transport Merge

• precision: dtype: bfloat16

• Context length: 262,144

Parameter Settings:

General Tasks:

temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Reasoning Tasks:

temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0

Paper & GitHub Repository:

Paper

GitHub Repository

Some Details:

Data:

CodeAlpaca_20K

We perform activation extraction using 2,000 prompts from the test dataset.

Details of Each Tensor:

merge_stats.csv

global_state.json