steven0226/qwen3vl-8b-chartqa-merged-16bit

VISIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The steven0226/qwen3vl-8b-chartqa-merged-16bit model is an 8 billion parameter Qwen3-VL architecture, fine-tuned by steven0226 for ChartQA tasks. This 16-bit merged checkpoint serves as a quality reference and serving baseline, achieving an 86.24% overall accuracy on the ChartQA test set. It is specifically optimized for visual question answering on charts and graphs, demonstrating strong performance in extracting and interpreting data from visual representations.

Loading preview...

Overview

This model, steven0226/qwen3vl-8b-chartqa-merged-16bit, is a full merged checkpoint derived from fine-tuning the Qwen3-VL-8B base model with a ChartQA LoRA adapter. It represents the 16-bit quality reference before AWQ quantization, offering a robust baseline for performance. The model was trained on 15,000 shuffled ChartQA examples over one epoch, with a maximum sequence length of 2,048 tokens. LoRA adaptation was applied to vision, language, attention, and MLP modules.

Key Capabilities

  • Chart-based Visual Question Answering (ChartQA): Achieves an overall accuracy of 86.24% on the ChartQA test set (77.28% on Human split, 95.20% on Augmented split) with a 5% numeric tolerance.
  • Multimodal Processing: Leverages the Qwen3-VL architecture for integrated vision and language understanding.
  • Serving Baseline: Provides a 16-bit serving reference with measured performance metrics, including output tokens per second and end-to-end latency, for various concurrency levels.

Usage Considerations

  • Primary Use Case: Best suited for tasks involving question answering based on visual charts and graphs.
  • Limitations: Primarily evaluated on English ChartQA short answers and not validated for general OCR documents, out-of-domain charts, or safety-critical applications.
  • Deployment: Can be loaded using a Qwen3-VL-compatible Transformers/vLLM environment. For serving, the downstream AWQ checkpoint is recommended if a minor quality trade-off (-0.72 pp) is acceptable for reduced size and potentially faster inference.

Finetunes
1 models