DeepSeek is back! JanusFlow 1.3B – Unified multimodal LLM > Key Finding: Rectified flow can be trained within the large language model framework without complex modifications. > Base Model: Built on DeepSeek-LLM-1.3b-base. > Vision Encoder: SigLIP-L, supports 384 x 384

→ View original post on X — @reach_vb,