GRACE-VLM

Live demo for GRACE-VLM: INT4 Quantization-Aware Distillation for Vision-Language Models, accepted at ICML 2026. Read the paper at arXiv:2601.22709.

Upload an image and ask a question to run the GRACE 2B BF16 checkpoint on free ZeroGPU hardware. For genuine packed INT4 inference, use ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ with the copy-ready GRACE loader.