HWBJTUOPD/Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

HWBJTUOPD/Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher is a 9 billion parameter vision-capable language model, developed by HWBJTUOPD, based on the Qwen3.5 architecture. This model is an OPD student, fine-tuned with 65 Vision-OPD update steps from an SFT checkpoint, distinguishing it from its base. It excels in visual understanding tasks, achieving 92.15% on VStar and 74.4955% macro score across four benchmarks, making it suitable for applications requiring robust multimodal reasoning.

Loading preview...

Model Overview

HWBJTUOPD/Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher is a 9 billion parameter vision-capable model, developed by HWBJTUOPD, derived from the Qwen3.5-9B architecture. This model is specifically an "OPD student" that underwent 65 Vision-OPD update steps, starting from a 10K SFT checkpoint. It was trained using a fixed Qwen3.5-9B teacher, focusing on visual understanding tasks.

Key Capabilities & Performance

This model demonstrates strong performance in multimodal benchmarks:

  • VStar: Achieves 176/191 (92.15%) on the public VStar evaluation, which is reproducible with provided scripts.
  • Macro Score: Attains a 74.4955% macro score across four benchmarks: VStar, MMStar (77.27%), BLINK-v5 (66.44%), and ZoomBench (62.13%).
  • Vision-OPD Training: Utilizes the yuanqianhao/Vision-OPD-6K dataset (6,241 rows) and the Vision-OPD runtime for its specialized training, where the student views full images with marked regions and the teacher views cropped target regions.

When to Use This Model

This model is particularly well-suited for:

  • Vision-Language Tasks: Applications requiring robust understanding and reasoning over visual inputs, especially those involving object detection or region-specific analysis.
  • Research & Reproduction: Developers interested in replicating or extending the Vision-OPD methodology, as the repository provides detailed configurations and evaluation scripts.
  • High-Performance Multimodal AI: For use cases demanding strong benchmark performance in combined vision and language understanding, as indicated by its competitive macro score.

Finetunes
1 models