Nvidia has released SANA-WM, a 2.6-billion-parameter open-source world model that generates up to 60 seconds of 720p video with precise camera control on a single GPU. The model, published as an arXiv preprint on May 14 with weights following days later, is the first open-source system of its kind to approach the quality bar set by closed rivals such as OpenAI’s Sora. It is built explicitly for the use case that matters most to Nvidia: simulating the physical world for robots and autonomous systems.
What SANA-WM does
SANA-WM is trained natively for minute-scale generation – not stitched together from short clips, but trained to produce a coherent 60-second video in one pass. It outputs 720p footage with 6-DoF camera control, meaning the camera trajectory can be specified precisely in metric-scale coordinates, a requirement for generating the kind of viewpoint-consistent data that robotics models need. The system was trained on roughly 213,000 public video clips with metric-scale pose annotations and completed training in 15 days on 64 H100 GPUs – a modest budget by frontier standards.
At inference, the model runs on a single GPU. Three variants are offered: a bidirectional generator for high-quality offline synthesis, a chunk-causal autoregressive generator for sequential rollout, and a distilled few-step variant that can denoise a full 60-second 720p clip in 34 seconds on a single RTX 5090 using NVFP4 quantization. Nvidia released the model under the Apache 2.0 license, making it free for commercial use and modification. The system is built on the SANA-Video codebase, and the streaming variant streams all weights on first use, so researchers can run it without manual downloads.
The quality numbers are competitive with far larger systems. On Nvidia’s one-minute world-model benchmark, SANA-WM demonstrates stronger action-following accuracy than prior open-source baselines, and it matches the visual quality of industrial-scale systems at roughly 36 times their throughput. With the optional 17-billion-parameter refiner active, camera rotation error drops from 7.59 to 4.50 degrees and camera motion consistency improves from 1.63 to 1.41, while the base system produces about 24 clips per hour on a single GPU. The efficiency gap versus rivals is stark: an 8-billion-parameter system like HY-WorldPlay manages roughly 1.1 clips per hour and needs eight GPUs to do it.
How it is built
The architecture is where SANA-WM departs from earlier video diffusion models. Generating 60 seconds of 720p video means processing nearly a thousand latent frames, a sequence length that makes standard softmax attention prohibitively expensive. Nvidia’s solution is a hybrid linear attention scheme that interleaves 15 frame-wise Gated DeltaNet blocks with 5 softmax attention blocks across 20 total transformer layers. The linear blocks keep memory and compute costs growing at a manageable rate, while the sparse softmax blocks preserve exact long-range recall where recurrence alone is insufficient.
A two-stage pipeline handles final quality: the main generator produces the minute-long output, and a dedicated refiner then corrects structural artifacts and sharpens details across the full duration. Training data combines real footage and synthetic renderings, including 3D Gaussian splatting renders and game-engine captures, all carrying metric-scale camera pose annotations that teach the model the relationship between camera motion and scene geometry. A causal LTX-2 VAE decodes the latent output chunk by chunk, and a Gemma text encoder handles prompts.
Why it matters
The significance of SANA-WM goes beyond video generation. World models – systems that learn the dynamics of a physical scene and predict what happens next – are the technology that robotics and embodied AI companies have been waiting for. A robot arm learning to pick objects needs thousands of clips from specific angles with specific camera motions; generating those synthetically with precise control is dramatically cheaper than collecting them in the real world.
The release also changes the structure of the video generation market, moving from the “make it move” stage to a “make it understand the world” stage. Sora demonstrated what closed systems could do; SANA-WM shows that an open model can reach comparable capability while running on hardware that fits in a standard server rack. For researchers without access to enormous clusters, that is the difference between studying world models and building on them. And because the model is Apache 2.0 licensed, it can be folded directly into commercial robotics pipelines without legal wrangling, which is precisely the ecosystem effect Nvidia wants.
Nvidia has done what it does best: turned a frontier research result into an open, accessible tool that strengthens the ecosystem around its hardware. SANA-WM is not just a video model – it is a simulation engine for the physical world, released under a permissive license, and it lowers the barrier to entry for every robotics lab that needs synthetic training data. The question is how quickly the industry’s closed labs respond.
