GitHub repository for real-time long video generation

LongLive 2.0 is an NVFP4 parallel infrastructure for long video generation, designed for both training and inference to enable real-time processing. It was released on May 13, 2026, and the existing LongLive 1.0 was split off into the v1.0 branch.

On the training side, the key features are balanced sequence parallel, multi-shot/single-shot AR training, and support for NVFP4 or BF16. For inference, it provides NVFP4 inference (W4A4), NVFP4 KV cache, multi-shot attention sink, sequence parallel inference, and async decoding.

Results were also presented. The LongLive-2.0-5B NVFP4 2-step model achieved 45.7 FPS and recorded VBench 83.14. The models released together are LongLive-1.3B, LongLive-2.0-5B, LongLive-2.0-5B-NVFP4-4Step, and LongLive-2.0-5B-NVFP4-2Step, with documentation, demos, and training/inference code organized on GitHub and Hugging Face.

Also, version 1.0 focused on real-time interactive long video generation, presenting attention sink, KV-recache, and streaming long tuning as its core ideas. Version 2.0 adds parallel infrastructure and low-precision inference on top of this, boosting the throughput and real-time performance of long video generation.