Mage-VL

Name: Mage-VL
Author: Microsoft

Microsoft · 4.7B parameters

Mage-VL is Microsoft's open-weight multimodal generative checkpoint for image and video understanding. It combines the Mage-ViT visual encoder, a two-layer projector, and the Qwen3-4B-Instruct-2507 language backbone, with a cognition gate for event-triggered streaming responses. The complete BF16 checkpoint has 4,741,793,792 parameters under Apache 2.0; its language configuration supports 262,144 positions and its tokenizer assigns 151,665 IDs.

Benchmark scores

Benchmark	Score
Roboflow Vision Evals — Counting, Judged (September)	26.6
Roboflow Vision Evals — Counting, Strict (September)	26.6
Roboflow Vision Evals — Detection mAP@50 (September)	2.5
Roboflow Vision Evals — Detection mAP@50:95 (September)	0.7
Roboflow Vision Evals — Detection mAP@75 (September)	0.0
Roboflow Vision Evals — Extraction, Judged (September)	55.0
Roboflow Vision Evals — Extraction, Strict (September)	47.1
Roboflow Vision Evals — Identification, Judged (September)	59.4
Roboflow Vision Evals — Identification, Strict (September)	44.8
Roboflow Vision Evals — OCR (September)	60.6
Roboflow Vision Evals — Reasoning, Judged (September)	24.7
Roboflow Vision Evals — Reasoning, Strict (September)	18.8
Roboflow Vision Evals (October 2026, four tasks without OCR)	28.3
Roboflow Vision Evals (September 2026 protocol)	38.1