OSP-Next: Efficient High-Quality Video Generation with Sparse Sequence Parallelism, HiF8 Quantization, and Reinforcement Learning

Researchers introduced OSP-Next, a text-to-video model designed to cut compute and communication costs while keeping video quality high.

It combines hybrid full-sparse attention, Skiparse-2D Attention, Sparse Sequence Parallelism, HiF8 quantization, and Mix-GRPO post-training.

The model outperforms a baseline on VBench and runs faster on NVIDIA H200 and Ascend hardware.

The quantized version shows only a small quality drop, suggesting a practical path to cheaper high-resolution video generation.