Mage-Flow

Microsoft's Mage project holds every model to a fixed 4B parameter budget, and Mage-Flow is its text-to-image and instruction-based editing member, pairing the Mage-VAE tokenizer with a native-resolution multimodal diffusion transformer, NR-MMDiT, trained as a rectified flow. Despite the small budget, the RL-aligned release posts 0.90 on GenEval, ahead of open models five to eight times larger, including FLUX.2-dev at 32B and Qwen-Image at 20B, which both score 0.87. A distilled 4-step Turbo variant renders a 1024px image in about 0.6 seconds on a single A100, and dedicated editing checkpoints handle instruction-driven modification of existing images. Base, RL-aligned, Turbo, and editing weights all sit on Hugging Face under the microsoft organization, released in July 2026 alongside code, with MIT licensing on Mage-Flow leaving commercial use unrestricted. The compact size keeps inference and fine-tuning within a single workstation-class GPU, exactly the niche it targets against far heavier open image models.