SGLang

VisionWeave Creates Elastic Visual Representations as a Native Capability of Multimodal Large Language Models

VisionWeave uses elastic visual representations, reducing visual tokens 43.0% on Qwen3.8-27B while retaining 9…