Microsoft Unveils Mage-VL: Codec-Native Streaming VLM with 3.5x Inference Speedup

pmttyji · reddit · 2026-07-29

Microsoft introduced Mage-VL, a 4B parameter codec-native streaming multimodal foundation model designed for efficient image and video understanding.

Core Innovation: The visual encoder is trained entirely from scratch. Borrowing the I/P frame mechanism from video codecs, it cuts visual tokens by over 75% by keeping anchor frames and only extracting patches from regions with actual motion.
Inference Speedup: Achieves up to 3.5x wall-clock inference speedup over uniform frame sampling at matched accuracy, featuring native-resolution scaling.
Dual-Process Design: Implements a System 1 (a lightweight cognition gate for low-latency monitoring) and System 2 (the full VLM for complex events) within a single model to enable proactive streaming.
Performance: With a fixed Qwen3-4B LLM backbone, it significantly outperforms the similarly-sized Qwen3-VL-4B across video understanding and spatial intelligence benchmarks.

More from Models