TencentBAC/TBAC-VLR1-3B-preview
TBAC-VLR1-3B-preview is a 3 billion parameter multimodal language model developed by Tencent PCG Basic Algorithm Center. Based on Qwen2.5-VL-3B-Instruct, it is fine-tuned using Group Relative Policy Optimization (GRPO) to significantly enhance multimodal reasoning capabilities. The model excels in complex reasoning tasks across various benchmarks, making it suitable for applications requiring advanced visual and mathematical understanding.
Loading preview...
Overview
TBAC-VLR1-3B-preview is a 3 billion parameter multimodal language model developed by the Tencent PCG Basic Algorithm Center. It is built upon the Qwen2.5-VL-3B-Instruct architecture and utilizes Group Relative Policy Optimization (GRPO) to significantly improve its multimodal reasoning abilities. This model demonstrates strong performance on several multimodal reasoning benchmarks, positioning it as a competitive option in its size class.
Key Capabilities
- Enhanced Multimodal Reasoning: Specialized fine-tuning with GRPO allows the model to perform complex reasoning tasks involving both visual and textual information.
- Strong Benchmark Performance: Achieves an average score of 35.7 across key multimodal reasoning benchmarks such as MathVista, MathVision, MathVerse, DynaMath, WeMath, and LogicVista, outperforming other 2B-3B parameter models like Qwen2-VL-2B, InternVL2.5-2B, and Qwen2.5-VL-3B.
- Instruction Following: Designed to follow instructions for problem-solving, as indicated by its system prompt for a helpful assistant that provides reasoning processes and answers.
Good for
- Complex Visual Question Answering: Ideal for applications that require understanding and reasoning over images combined with textual queries.
- Mathematical Problem Solving: Particularly strong in benchmarks related to mathematical reasoning (e.g., MathVista, DynaMath, WeMath), suggesting suitability for tasks involving numerical and logical deduction from multimodal inputs.
- Research and Development: Useful for researchers and developers exploring advanced multimodal AI, especially those focused on improving reasoning capabilities in smaller models.