LingBot-Video Open-Sourced

LingBot-Video is a single-stream Diffusion Transformer model that adopts a DeepSeek-V3-style Sparse MoE architecture. It has 13B total parameters, but uses only 1.4B active parameters during inference.

Key features are as follows:

• RL Post-training: Reinforcement learning was performed using a 6-component reward scheme, including physical-plausibility rewards.
• Action-Video Mode: It predicts robot motion in video form based on the robot's Action and Hand-pose conditions.
• Performance: It achieved the best average performance on RBench, and also demonstrates excellent performance in terms of video generation quality.

The model weights, code, and Diffusers/SGLang stack are currently open-sourced.