Microsoft Releases Mage-VL: An Efficient Codec-Native Streaming Multimodal Model

Microsoft has released Mage-VL, a 4-billion parameter multimodal foundation model that leverages video codec principles to process streaming video. By focusing only on areas of motion and new details (P-frames) alongside anchor frames (I-frames), it reduces visual tokens by over 75% and speeds up inference by up to 3.5x. Traditional video LLMs are computationally heavy and slow for real-time streaming because they process dense grids of uniform frames. Mage-VL addresses this bottleneck, enabling efficient, low-latency real-time video understanding and event-driven commentary on consumer-grade hardware. The model pairs a from-scratch visual encoder (Mage-ViT) with a Qwen3-4B language decoder, utilizing a dual-process "System 1 & System 2" design where a lightweight cognition gate filters out routine content before invoking the full VLM. It is codec-agnostic, supporting traditional codecs like H.264 and HEVC as well as neural codecs like DCVC-RT without architectural changes.