Kingsoft-LLM/QZhou-Flowchart-VL-32B

VISIONPricing:Input $1.06 / Cached $0.053 / Output $2.6Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kPublished:Nov 19, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

QZhou-Flowchart-VL-32B by Kingsoft-LLM is a 32 billion parameter multimodal large language model specifically designed for understanding and reasoning about flowcharts. Post-trained from Qwen2.5-VL-32B using Group Relative Policy Optimization (GRPO) and a novel Tri-CoT (Three-stage Chain-of-Thought) reasoning structure, it excels at structured analysis of flowchart diagrams. The model achieves 87.83% on the QZhou-Flowchart-QA-Benchmark, outperforming larger models like GPT-5 and Gemini-2.5-Pro in this specialized domain.

Loading preview...

Overview

Kingsoft-LLM's QZhou-Flowchart-VL-32B is a 32 billion parameter multimodal large language model (MLLM) specialized in flowchart understanding and reasoning. Built upon Qwen2.5-VL-32B, this model leverages a unique Tri-CoT (Three-stage Chain-of-Thought) reasoning structure and is fine-tuned using Group Relative Policy Optimization (GRPO) reinforcement learning.

Key Capabilities & Features

  • Specialized Flowchart Reasoning: Designed for accurate interpretation and logical inference from flowchart diagrams.
  • Tri-CoT Structured Reasoning: Employs a three-stage process: JSON extraction of flowchart structure, logical inference, and final answer generation, ensuring grounded reasoning.
  • High Performance: Achieves 87.83% on the QZhou-Flowchart-QA-Benchmark, surpassing models like GPT-5 and Gemini-2.5-Pro in this specific task.
  • Reinforcement Learning: Utilizes GRPO for post-training, which enhances structured reasoning and reduces hallucinations while maintaining general multimodal capabilities.
  • Robust Training: Trained on 14,600 synthetic flowchart samples with Tri-CoT annotations, focusing on diverse question types including node identification, path analysis, and spatial relationships.

When to Use This Model

QZhou-Flowchart-VL-32B is ideal for applications requiring precise and structured understanding of flowcharts. Its ability to extract detailed JSON representations of diagrams before reasoning makes it suitable for automated analysis, documentation, and question-answering systems where flowchart interpretation is critical. While maintaining strong general multimodal capabilities, its primary strength lies in its specialized performance on flowchart-related tasks.

Limitations

  • Primarily trained on synthetic flowcharts; performance on real-world, noisy images may require further adaptation.
  • May experience degraded performance with extremely large flowcharts (e.g., over 50 nodes).

Finetunes
1 models