Back

DeepSeek just dropped open-source Janus Pro 7B for image understanding and generation! — SOTA 0.8 on GenEval and 84.19 on DPG-Bench, beats DallE3 and SD3-Medium — 72M synthetic images in pretraining — good text rendering Images are small (384x384)

how are they able to even do this so quickly??

I believe they were silently training the model while the west was trying to capture the market.