Mage (GitHub Repository)

Mage is a lightweight, fixed-budget multimodal model family with 4B parameters released by Microsoft, consisting of two models that handle image/video understanding and image generation/editing respectively.

Mage-VL is a codec-native vision-language model for image/video understanding and proactive streaming, trained from scratch at the 4B scale. Code and checkpoints are planned for release.

Mage-Flow is a generative model for text-to-image generation and instruction-based image editing, composed of Mage-VAE (a lightweight, high-fidelity latent tokenizer) and a Native-Resolution Multimodal Diffusion Transformer. It is provided in three variants: Base, RL-aligned, and 4-step Turbo.

Key features:

• Efficient tokenizer: Mage-VAE reduces encoding/decoding compute by ~12×/~22× respectively compared to FLUX.2-VAE
• Native resolution: A single checkpoint supports 512–2048, up to extreme 4:1 aspect ratios
• Training speed: Training time per step improved from ~1.93s to ~0.78s (~2.5× improvement)
• Inference speed: On a single A100 GPU, generating a 1024² image takes 0.59 seconds, editing takes 1.02 seconds
• Achieves equal or superior performance at a much smaller scale than competing models (Qwen-Image 20B, FLUX.2 32B, etc.)