V-Retrver/V-Retrver-SFT-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Feb 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

V-Retrver/V-Retrver-SFT-7B is a 7 billion parameter multimodal large language model (MLLM) developed by V-Retrver, designed for universal multimodal retrieval. It employs an evidence-driven agentic reasoning framework, allowing it to actively verify fine-grained visual evidence through external visual tools. This model excels at multimodal interleaved reasoning, alternating between hypothesis generation and targeted visual verification, leading to significant improvements in retrieval accuracy and perception-driven reasoning reliability.

Loading preview...

V-Retrver-SFT-7B: Evidence-Driven Multimodal Retrieval

V-Retrver-SFT-7B is a 7 billion parameter multimodal large language model (MLLM) specifically designed for universal multimodal retrieval tasks. Developed by V-Retrver, this model introduces an innovative evidence-driven agentic reasoning framework, as detailed in its accompanying paper. Unlike traditional language-driven approaches that rely on static visual encodings, V-Retrver-SFT-7B actively engages in visual inspection to verify evidence, addressing issues of speculative reasoning in visually ambiguous scenarios.

Key Capabilities

  • Evidence-Driven Agentic Reasoning: Reformulates multimodal retrieval as an agentic process grounded in visual inspection.
  • Multimodal Interleaved Reasoning: Enables the MLLM to selectively acquire visual evidence using external visual tools, alternating between hypothesis generation and targeted visual verification.
  • Enhanced Retrieval Accuracy: Demonstrates consistent improvements in retrieval accuracy, averaging 23.0% across multiple multimodal retrieval benchmarks.
  • Perception-Driven Reasoning Reliability: Improves the reliability of reasoning by actively verifying visual evidence.
  • Generalization: Shows strong generalization capabilities across various multimodal retrieval tasks.

Good for

  • Applications requiring robust universal multimodal retrieval.
  • Scenarios where fine-grained visual evidence verification is crucial.
  • Tasks benefiting from an agentic reasoning approach that integrates visual tools.
  • Research and development in multimodal AI, particularly in improving reasoning and retrieval accuracy.

Finetunes
1 models