Mage-VL cuts visual tokens by 75% while matching larger models in speed

Standard vision-language models suffer from Moravec's paradox: they excel at complex offline reasoning but fail at simple streaming perception. Mage-VL solves this with a bio-inspired dual-system architecture: a lightweight System 1 event gate for proactive streaming and a causal System 2 decoder. Its core innovation is Mage-ViT, a tokenizer that replaces uniform frame sampling with entropy-driven selection at the 16x16 patch level, using motion vectors and residual energy from sparse anchor (I) and predicted (P) frames. This cuts visual token consumption by over 75% while preserving spatiotemporal context. The model was trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, matching or beating flagship encoders trained on billions of image-text pairs.

The 4B-parameter Mage-VL matches Qwen3-VL-4B on static vision tasks and achieves strong gains in video understanding and 2D/3D spatial reasoning. It comprehensively surpasses the 15B-parameter Phi-4-reasoning-vision baseline, with up to 3.5x wall-clock inference speedup. The paper also introduces AI4AI data pipelines for prompt-code joint optimization and performance diagnosis, plus seven empirical findings on pre-training data efficiency, variable-resolution scaling, and motion-spatial synergy. This makes high-quality real-time multimodal understanding practical for streaming video without massive compute budgets.