Mage-VL
modelYour notes
"Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model" — a 4B VLM that allocates visual tokens like codec bits: >75% fewer visual tokens and up to 3.5× speedup while beating Qwen3-VL-4B on all reported video benchmarks, with an event-gated live-commentary mode. The Mage-ViT encoder is trained from scratch; the language decoder is Qwen3-4B. Apache 2.0; 435K downloads in its first ten days.
Mage-VL-Cowork (October 30, 2026, Apache 2.0) turns an intermediate 4B Mage-VL checkpoint into a computer-use agent that chooses, turn by turn, between clicking in the GUI and running code in a persistent shell, inside one policy with no external orchestrator. Before collecting teacher trajectories, Microsoft had an authoring agent prune and re-equip the teacher's harness, then synthesized tasks per atomic skill with its Mage-CUA-Forge engine. Training is a five-stage specialize-then-unify curriculum ending in online RL, about 7.79 ZFLOPs on 32 B200s plus 1,120 CPU sandboxes. It reports 66.0% on OSWorld-Verified at a 50-step budget, the best of any 4B model per the card. RL added 10.16 points over the hybrid SFT stage while cutting model calls by 13.0% and output tokens by 25.4%.