CodeGoat24/UnifiedReward-Flex-qwen35-9b

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 16, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

CodeGoat24/UnifiedReward-Flex-qwen35-9b is a 9 billion parameter model designed as a unified personalized reward model for vision generation. Based on the Qwen architecture, it couples reward modeling with flexible and context-adaptive reasoning. This model is specifically optimized for evaluating and guiding vision generation tasks, aiming to mitigate position bias through updated weights and enhanced training data.

Loading preview...

UnifiedReward-Flex-qwen35-9b: Vision Generation Reward Model

UnifiedReward-Flex-qwen35-9b is a 9 billion parameter model developed by CodeGoat24, specifically engineered as a unified personalized reward model for vision generation. This model integrates reward modeling with flexible and context-adaptive reasoning capabilities, making it suitable for evaluating and guiding visual content creation.

Key Capabilities & Features

  • Personalized Reward Modeling: Designed to provide personalized feedback for vision generation tasks.
  • Flexible & Context-Adaptive Reasoning: Incorporates reasoning that adapts to varying contexts within vision generation.
  • Mitigated Position Bias: Features updated model weights and enhanced training data to address and reduce position bias issues.
  • Vision Generation Focus: Primarily aimed at improving the quality and relevance of generated visual content.

Use Cases

  • Evaluating Generated Images: Can be used to assess the quality and alignment of images produced by other generative models.
  • Guiding Generative AI: Provides reward signals to steer the training or inference of vision generation models towards desired outputs.
  • Research in Vision-Language Models: Useful for researchers exploring personalized feedback mechanisms in multimodal AI.

For more technical details, refer to the associated paper and the project page.