Kimhi/AWARES-Qwen2.5-VL-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:May 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Kimhi/AWARES-Qwen2.5-VL-7B is a 7 billion parameter vision-language model, fine-tuned from Qwen2.5-VL-7B-Instruct using GRPO. This model features an adaptive resolution mechanism, allowing it to intelligently request high-resolution image crops when needed to answer visual questions. Its primary differentiator is the ability to dynamically zoom into specific image regions, making it highly efficient for tasks requiring detailed visual analysis without processing the entire image at high resolution.

Loading preview...

AWARES-Qwen2.5-VL-7B: Adaptive Resolution for Efficient VLMs

Kimhi/AWARES-Qwen2.5-VL-7B is a 7 billion parameter vision-language model, building upon the Qwen2.5-VL-7B-Instruct architecture. Its core innovation lies in Adaptive Resolution with Active REqueStingS (AWARES), a mechanism that allows the model to dynamically request high-resolution image crops only when necessary to answer visual questions. This approach significantly enhances efficiency by avoiding the processing of entire images at high resolution.

Key Capabilities

  • Intelligent Crop Requesting: Given a low-resolution image and a question, the model determines if more visual detail is required and emits a GET_CROPS tool call for specific regions.
  • Dynamic Zoom: Supports requesting any of 9 predefined crop indices (e.g., top-left, center, bottom-right) to focus on relevant areas.
  • Efficient Visual Question Answering: Optimizes resource usage by processing high-resolution data only for critical image segments.
  • Multi-turn Inference: Designed to handle a multi-turn pipeline where high-res crops are provided back to the model for final answer generation.

Good For

  • Applications requiring detailed visual analysis where full high-resolution processing is computationally expensive.
  • Tasks where specific objects or regions within an image are crucial for accurate answers.
  • Research and development in adaptive vision-language models and efficient multimodal AI.
  • Scenarios benefiting from tool-use capabilities for dynamic data retrieval.

Finetunes
1 models