NVIDIA released Long-WAM for long-context world-action models. It supports robot policy training, evaluation, and real-robot deployment. LongLive-Plug enables once-for-all distillation for video generation. LongLive 2.0 provides NVFP4 parallel infrastructure for long video generation. LongLive 1.0 allows real-time interactive long video generation. Each project is self-contained with its own code, documentation, and model weights. Demo videos are available on YouTube. Model checkpoints are hosted on Hugging Face. The Long-WAM paper is now on arXiv. LongLive 2.0 supports FP8 inference. LongLive 2.0 achieved 45.7 FPS on a GB200. NVIDIA optimized the NVFP4 inference path for an 18.6% throughput gain.
What it unlocks
Generating long videos in real time with streaming visual memory for robot policies. Distilling video generation capabilities once for reuse across models. Training and serving long-video models faster with NVFP4 quantization and sequence parallelism.
- 45.7 FPS on GB200
- 18.6% throughput gain
- 50% KV cache reduction
- 60s real-time interactive videos
What you need to act on it
- access to NVIDIA hardware (e.g., GB200, RTX 5090, DGX Spark, AGX Thor)
- technical expertise in model training and deployment
Sources