NVlabs has released SANA-Video 2.0, a video diffusion model available in two scales (5B and 14B parameters) that generates video up to 720p on a single GPU. The model uses a hybrid architecture combining linear attention and softmax in calibrated proportions (3:1), maintaining the quality of traditional video transformers while with computational cost that scales linearly. On paper, the 5B model produces 720p video from 5 seconds in 13 seconds on an H100, making it 120 times faster than Wan 2.2-A14B on the same hardware (paper published July 23, 2026 on Hugging Face).

Why it matters to you

If you generate AI video content, cost has been the bottleneck so far. Quality video models run in the cloud, with bills climbing quickly for every second of footage. A model that produces 720p on a single GPU changes the equation: instead of paying an API for each clip, you can keep the process local, iterate without watching the meter, and control what goes in and what comes out.

The code is on GitHub (NVlabs/Sana), so anyone with a sufficiently powerful GPU can try it directly.

If you want to try it

The repo is at github.com/NVlabs/Sana. Before investing time, check how much video memory the 5B model requires and compare it to your card: the performance numbers are measured on H100, a data center GPU, not a consumer card.

In detail

The problem with video models is sequence length. A video is a long chain of frames, and each frame consists of thousands of tokens. The standard mechanism (softmax attention) compares every token with every other, so the cost grows quadratically: double the video length and the cost quadruples. That’s why generating 10 seconds of video costs much more than 10 times generating 1 second.

Linear attention solves the cost problem but loses some expressiveness: it can’t capture certain fine interactions between distant tokens. SANA-Video 2.0 uses a compromise: for most layers it uses linear attention (linear cost), but periodically inserts softmax pillars (one every three linear layers) that restore the ability to capture complete interactions. The 3:1 ratio was chosen after testing on reduced-resolution proxies, not intuition.

The second technical trick is Block Attention Residuals: instead of recalculating rich representations at each softmax pillar, the model routes them through the subsequent linear layers, reusing work already done. This increases the effective rank of deep layers by 12%, according to the authors.

All of this is trained from scratch, not a simplification of an existing model. The difference matters: linearizing an already-trained model is cheaper but degrades quality. Training the hybrid architecture from scratch costs more in training but produces a model that knows how to work with its structure from the start.

The speed numbers are measured on H100, a data center GPU that costs as much as a car. The jump to consumer GPU (RTX series 40 or 50 that a professional might have on their desktop) is neither immediate nor guaranteed. End-to-end optimizations (kernel fusion, caching, sparse attention) are tuned for specific hardware. Before considering it feasible on your own machine, you need to see real numbers on an RTX 4090 or 5090.

The comparison with Wan 2.2-A14B (120 times slower on the same H100) is suggestive but partial: Wan is a larger model and the comparison is on SANA’s 5B model. SANA’s 14B model would have different numbers.

The code is open on GitHub, but full weights and setup instructions should be verified on the repo. The story is promising for those generating video content who want to escape the cloud pay-per-clip logic, but the transition from H100 in the lab to consumer GPU still has steps to climb.