UCSC-VLAA/VLAA-Thinker-Qwen2.5VL-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Mar 31, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

UCSC-VLAA/VLAA-Thinker-Qwen2.5VL-3B is a 3 billion parameter vision-language model developed by UCSC-VLAA, based on the Qwen2.5VL architecture. This model is specifically optimized for complex multimodal reasoning tasks, achieving state-of-the-art performance on the OpenCompass Multimodal Reasoning Leaderboard as of April 7th, 2025. It leverages a unique data generation pipeline, VLAA-Thinking, which includes sophisticated visual captioning, reasoning answer distillation using DeepSeek-R1, and automated verification, making it highly effective for tasks requiring detailed visual and textual understanding.

Loading preview...

VLAA-Thinker-Qwen2.5VL-3B Overview

VLAA-Thinker-Qwen2.5VL-3B is a 3 billion parameter vision-language model from UCSC-VLAA, built upon the Qwen2.5VL architecture. It is designed for advanced multimodal reasoning, demonstrating state-of-the-art performance on the OpenCompass Multimodal Reasoning Leaderboard as of April 7th, 2025. This model's strength lies in its sophisticated training methodology, which includes a novel data generation pipeline called VLAA-Thinking.

Key Capabilities

  • Advanced Multimodal Reasoning: Excels at tasks requiring deep understanding of both visual and textual information.
  • State-of-the-Art Performance: Achieves top results on the OpenCompass Multimodal Reasoning Leaderboard.
  • Robust Data Generation: Utilizes a unique VLAA-Thinking pipeline for high-quality training data, involving:
    • Metadata Collection: Gathers data from 9 diverse vision-language datasets.
    • Visual Captioning: Generates detailed, structured captions using GPT-4o.
    • Reasoning Answer Distillation: Employs DeepSeek-R1 to produce structured, logical reasoning steps.
    • Answer Refinement & Verification: Uses GPT-3.5-turbo for clarity and automated validation against ground-truth answers.
  • Specialized Training Data: Curates distinct Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) datasets, with RL focusing on more challenging reasoning examples containing "aha moments."

Good For

  • Complex Visual Question Answering (VQA): Ideal for scenarios demanding intricate reasoning over images and associated questions, such as those found in CLEVR-Math, GeoQA, and ArxivQA.
  • Research in Multimodal AI: Provides a strong baseline and framework for investigating reasoning capabilities in vision-language models, particularly concerning SFT vs. RL approaches.
  • Applications Requiring High Accuracy in Reasoning: Suitable for tasks where precise, step-by-step logical deduction from visual and textual inputs is critical.

Finetunes
1 models