Microsoft's 4B multimodal family: understand, generate, and click

What it does

Mage is a family of three 4B-parameter models from Microsoft: Mage-VL for image and video understanding, Mage-Flow for text-to-image generation and instruction-based editing, and Mage-VL-Cowork for hybrid computer use mixing GUI clicks with code and terminal actions. The pitch is a fixed compute budget — small enough to fine-tune on modest hardware, competitive with much larger open models in each domain. Everything is on Hugging Face with per-model sub-READMEs.

The interesting bit

The unifying idea is “codec-aligned efficiency”: spend representation capacity where the signal actually is. On the understanding side, a from-scratch visual encoder (Mage-ViT) uses a bio-inspired I/P predictive-patch mechanism to cut visual tokens by over 75%, yielding up to 3.5× faster inference and 8× longer video training. On the generation side, Mage-VAE matches FLUX.2-VAE reconstruction fidelity with roughly 12–22× fewer encode/decode MACs per pixel, which removes the usual high-resolution VAE bottleneck.

Key highlights

• Mage-VL beats Qwen3-VL-4B on every reported video and temporal-grounding benchmark (e.g. +22.5 QVHighlight, +11.0 VSI-Bench) with the same 4B Qwen3 decoder.
• One Mage-VL checkpoint bundles understanding plus a proactive streaming gate for low-latency, event-gated commentary — no separate variants.
• Mage-VL-Cowork hits 66.0% on OSWorld-Verified at 4B under a 50-step budget, with online hybrid RL adding 10.16 points over SFT.
• Mage-Flow generates at native resolution from 512 to 2048 on any aspect ratio, including extreme 4:1; the Turbo variant does 1024² images in 0.59 s on a single A100.
• Mage-Flow-Edit handles semantic editing, appearance changes, restoration, and structure-aware outputs in one unified model.

Caveats

• The claims are benchmark-heavy and self-reported; the README doesn’t detail failure modes or what the models are bad at.
• It’s a research family, not a product — three separate models with their own training recipes, so don’t expect one drop-in multimodal assistant.
• The proactive streaming gate is demonstrated on sports broadcasts; generalization beyond that is left to the paper.

Verdict

If you do post-training research or need multimodal models you can actually retrain without a datacenter budget, this is squarely aimed at you. If you just want the biggest numbers on a leaderboard, the 4B ceiling is the whole point — look elsewhere.