LingBot-Video: sparse-MoE video diffusion transformer (13B total, 1.4B active) post-trained as an action-conditioned world model[R]

Single-stream diffusion transformer with a DeepSeek-V3-style sparse MoE (128 experts, top-8 routing, 1.4B active of 13B total). Six-reward RL post-training including a physical-plausibility reward, plus an action-to-video mode that predicts robot rollouts from action and hand-pose conditions. Weights, code, and a Diffusers/SGLang stack are open under the LingBot-Video name. Two things I would push on, and would genuinely like this sub's read: The physical-plausibility reward is graded by a VLM from sampled