PatchyTisa/Qwen3.6-35B-A3B-Caption-Eagle3DraftModel

Eagle3 Draft Model - Qwen3.6-35B-A3B Fine-tuned Version

Model Description

This is an Eagle3 (Enhanced Auxiliary Loss for Efficient Speculative Decoding) draft model derived from the open-source Qwen/Qwen3.6-35B-A3B base model.

Training Details

• Base Model: Qwen/Qwen3.6-35B-A3B
• Training Dataset: Microsoft COCO train2017 caption (100k samples)
• Training Method: Speculative decoding draft model training with Eagle3 architecture
• Average Acceptance Rate: 83%

Model Performance

The model achieves high speculative decoding efficiency with an average acceptance rate of 83%, significantly accelerating inference when used as a draft model in speculative decoding pipelines.

Usage

This model is designed to be used as a draft model for speculative decoding with the Qwen3.6-35B-A3B target model. It can be integrated with vLLM for efficient inference acceleration.

Files

• config.json - Model configuration file
• model.safetensors - Model weights

Citation

If you use this model, please consider citing the Eagle3 paper and Qwen3.6 model.