Microsoft unveils 4B image generation model Mage-Flow

Microsoft has unveiled Mage-Flow, which handles text-to-image generation and instruction-based editing in a single stack. Despite its 4B scale, it shows competitive quality against much larger models like Qwen-Image 20B and FLUX.2 32B.

The core components are two-fold:

• Mage-VAE: A lightweight tokenizer using 1-step diffusion encoding/decoding. Achieves ~12× reduction in encoding MACs and ~22× reduction in decoding MACs compared to FLUX.2-VAE
• NR-MMDiT: A native-resolution multimodal DiT that supports a single checkpoint across 512-2048 resolutions, up to extreme 4:1 aspect ratios

With native-resolution packing and fused CUDA kernels, training speed has been improved from ~1.93s to ~0.78s (a roughly 2.5x reduction).

Three variants are provided—Base, RL-aligned, and 4-step Turbo—and both Mage-Flow for generation and Mage-Flow-Edit for editing are publicly available on Hugging Face.