ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ
Image-Text-to-Text • 0.9B • Updated • 31 • 1
GRACE-VLM: Deployable INT4 Vision-Language Models
ICML 2026 · Qwen3-VL-2B distilled from 8B. Start with the real AWQ-packed INT4 checkpoint, retaining 98% of GRACE BF16.
Note Start here: real AWQ-packed INT4 checkpoint with a released Docker image.
GRACE-VLM
🦢1Try GRACE-VLM on an image, then deploy the INT4 build.
Note Live ZeroGPU demo: upload an image and try the GRACE 2B student.
ForeverBlue/Qwen3-VL-2B-GRACE-BF16
Image-Text-to-Text • 2B • Updated • 612 • 1Note Standard Transformers checkpoint and full-precision reference.
ForeverBlue/Qwen3-VL-2B-GRACE-W4G128
Image-Text-to-Text • 2B • Updated • 31 • 1Note W4G128 QAT research checkpoint for analysis and repacking.
ForeverBlue/Qwen3-VL-2B-GRACE-W8G128
Image-Text-to-Text • 2B • Updated • 34 • 1Note W8G128 QAT research checkpoint.
Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs
Paper • 2601.22709 • Published • 1Note GRACE paper, accepted at ICML 2026.
ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128-AWQ
Image-Text-to-Text • 1B • Updated • 14 • 1Note Real AWQ-packed LLaVA-1.5 deployment checkpoint.
ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128
Image-Text-to-Text • 7B • Updated • 11 • 1Note LLaVA-1.5 QAT checkpoint from the paper.
Lin-Chen/ShareGPT4V
Viewer • Updated • 1.35M • 2.1k • 315Note Training data used by the released GRACE models.