NVIDIA's SANA-Video 2.0: Hybrid Attention Model Runs 720p on a Single RTX 5090

mmowg · reddit · 2026-08-03

NVIDIA has released SANA-Video 2.0, a new Video Diffusion Transformer available in 5B and 14B parameter versions. It features a redesigned architecture with a 3:1 Hybrid Linear-Softmax Attention, combining the speed of linear attention with the expressiveness of softmax.

Powered by Sol-Engine optimization (kernel fusion, sparse attention, TensorRT, MXFP4/MXFP8), it achieves a 3.58x speedup. It is the first NVIDIA video model designed for consumer GPUs, capable of generating 720p video on a single RTX 5090. It scores 84.30 on VBench and is up to 120x faster than Wan 2.2-A14B on the same hardware.

Currently, the model is "research-open" with no code, weights, or licensing terms published yet.

Original post →

More from Multimodal

Multimodal channel →