Rex1090/PEARL-8B

VISIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Dec 5, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Rex1090/PEARL-8B is an 8 billion parameter multimodal reasoning model developed by Chi Zhang, fine-tuned from Qwen3-VL-8B-Instruct. This model specializes in Perceptual-Evidence Anchored Reinforced Learning for multimodal tasks, offering a 32768 token context length. It is designed for advanced multimodal reasoning applications, leveraging its specialized training for interpreting and processing diverse data types.

Loading preview...

Overview

Rex1090/PEARL-8B is an 8 billion parameter multimodal reasoning model developed by Chi Zhang, building upon the Qwen3-VL-8B-Instruct architecture. This model is specifically designed for Perceptual-Evidence Anchored Reinforced Learning (PEARL), focusing on enhancing multimodal reasoning capabilities. It supports a substantial context length of 32768 tokens, making it suitable for complex tasks requiring extensive input analysis.

Key Capabilities

  • Multimodal Reasoning: Excels at tasks that require understanding and integrating information from various modalities.
  • Perceptual-Evidence Anchored Learning: Utilizes a reinforced learning approach anchored by perceptual evidence, as detailed in its associated paper (arxiv.org/abs/2511.18437).
  • High Context Length: Benefits from a 32768-token context window, allowing for processing of longer and more detailed inputs.

Training Details

The model was fine-tuned using the ViRL39k dataset and leveraged the EasyR1 framework for its training process. Its development is documented in the PEARL GitHub repository.

Good For

  • Applications requiring advanced multimodal understanding.
  • Research and development in multimodal AI, particularly for reasoning tasks.
  • Use cases that can benefit from a model with a large context window for multimodal inputs.