openbmb/EVisRAG-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Oct 7, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

EVisRAG-3B by OpenBMB is a 3 billion parameter Vision Retrieval-augmented Generation (VisRAG) framework, built on Qwen2.5-VL-3B-Instruct, designed for multi-image questions. It enhances Visual Language Models (VLMs) by first linguistically observing retrieved images to collect per-image evidence, then reasoning over those cues to answer. The model is trained with Reward-Scoped GRPO, applying fine-grained token-level rewards to optimize both visual perception and reasoning, making it suitable for complex visual question answering tasks involving multiple images.

Loading preview...

EVisRAG-3B: Evidence-Guided Multi-Image Reasoning

EVisRAG-3B, developed by OpenBMB, is a 3 billion parameter Vision Retrieval-augmented Generation (VisRAG) framework built upon Qwen2.5-VL-3B-Instruct. This model is specifically engineered to address multi-image questions by equipping Visual Language Models (VLMs) with enhanced visual perception and reasoning capabilities.

Key Capabilities

  • Evidence-Guided Reasoning: EVisRAG processes multi-image questions by first linguistically observing retrieved images to gather per-image evidence, then performing reasoning based on these cues to formulate an answer.
  • Reward-Scoped GRPO Training: The model is trained using Reward-Scoped GRPO (RS-GRPO), a method that applies fine-grained token-level rewards. This joint optimization strategy improves both visual perception and reasoning accuracy.
  • End-to-End Framework: EVisRAG provides an end-to-end solution for multi-image scenarios, integrating precise visual perception directly into the reasoning process.

Good For

  • Complex Visual Question Answering: Ideal for tasks requiring the synthesis of information from multiple images to answer intricate questions.
  • Enhanced VLM Performance: Users looking to improve the visual perception and reasoning abilities of VLMs in multi-image contexts.
  • Research and Development: Provides a robust framework and methodology for further exploration in evidence-guided visual reasoning and retrieval-augmented generation.

Finetunes
1 models