You already have an account? Log In

NVIDIA's SANA-WM and the quiet rise of open-source AI video models

Posted on 16.05.2026

For most of the past two years, the most jaw-dropping AI video clips you've seen — the photorealistic mammoths, the cinematic dolly shots, the eerily believable street scenes — have come from closed systems. OpenAI's Sora, Google's Veo, Runway's Gen-models: all proprietary, all gated behind waitlists, credits, or enterprise contracts. If you're a creator in Brisbane or a researcher in Adelaide, you've largely been a spectator.

That picture is shifting. NVIDIA's release of SANA-WM, a 2.6-billion-parameter open-source "world model" capable of generating minute-scale 720p video on a single GPU, is the latest sign that the open ecosystem is closing the gap on the closed labs — not in raw spectacle, perhaps, but in the metrics that matter for everyday use: cost, control, and accessibility.

What SANA-WM actually is

SANA-WM is a video generation model from NVIDIA's research arm, released with weights and architecture details exposed for the community. The headline numbers are deliberately modest in a way that tells you who it's for: 2.6 billion parameters, 720p output, roughly a minute of generated footage, and crucially, it runs on a single GPU.

Compare that with the typical proprietary stack, which is usually backed by data-centre-scale inference and rented out per second of generated video. SANA-WM is engineered around the assumption that a single researcher, indie studio, or small VFX shop should be able to actually run the thing locally.

The "WM" — world model — part is also worth pausing on. A world model doesn't just predict the next frame of pixels; it tries to maintain an internal sense of physics, object permanence, and how a scene should evolve over time. That's the same conceptual lineage as Sora and Veo, but at a parameter count that's an order of magnitude smaller than what the closed competitors are believed to use.

Why the open/closed split matters for creators

If you make things for a living — short films, social content, ads, game prototypes, explainer animations — the difference between an open model and a proprietary one is more than ideological.

Cost predictability. Closed APIs charge per generation. An open model on your own hardware (or a rented cloud GPU) has a fixed cost. For an Australian creator already paying for a Veo or Runway subscription in USD, that matters.
Content policy control. Proprietary models enforce content rules at the API layer. That's often sensible, but it also means a music video director can't always generate, say, a stylised fight scene or a darkly-lit horror sequence. Open weights let users — and importantly, downstream platforms — set their own boundaries.
Fine-tuning. An open model can be adapted to a specific look: your studio's house style, a particular animation aesthetic, or a niche use case like architectural visualisation. That's effectively impossible with a closed API.
Local data. For Australian businesses dealing with privacy-sensitive footage — medical, legal, government — running inference on-prem is the only realistic path. SANA-WM's single-GPU requirement makes that genuinely viable.

The gap with proprietary tools — honestly assessed

It would be misleading to pretend SANA-WM matches Sora or Veo 3 frame-for-frame. The closed models still hold meaningful leads in a few areas:

Resolution and length

SANA-WM caps at 720p and around a minute. Veo 3 and Sora are pushing toward 1080p and beyond, with longer coherent sequences. For a TV commercial finished at 4K, that gap is real.

Prompt adherence and physics

The most expensive closed models have been trained on enormous, curated datasets and refined with extensive human feedback. They tend to handle complex prompts — "a glass of red wine tipping over onto a marble countertop in slow motion" — with more reliable physics. A 2.6B open model will be more brittle.

Audio and multimodal extras

Veo 3, for example, generates synchronised audio. SANA-WM is, at launch, a visual model. Expect the community to layer in audio pipelines, but it's another thing the closed players ship out of the box.

The trade-off, though, is the one that's defined open-source AI since Stable Diffusion in 2022: you give up some peak quality in exchange for total control, lower marginal cost, and the ability to actually build a product on top of the model rather than just consume it.

Single-GPU is the real headline

The most strategically important detail in SANA-WM isn't the resolution or the duration — it's that it runs on one GPU. That's the threshold at which technology stops being a research curiosity and becomes infrastructure.

Stable Diffusion crossed that threshold for images in 2022, and within months you had Photoshop plugins, indie apps, fine-tuned models for every conceivable aesthetic, and entire businesses built on top. Video has been waiting for its equivalent moment because the compute requirements have been brutal — minutes of GPU time for seconds of footage, and often on hardware most people can't access.

A 2.6B parameter model that fits on a single high-end consumer or prosumer GPU changes the maths. It's the difference between "interesting paper" and "something a 12-person Melbourne studio can deploy next quarter."

The infrastructure layer is catching up too

Open models are only half the story. To actually deploy AI video — or any generative AI — in production, you need the boring plumbing: session management, sandboxing, scaling, observability. That's a part of the stack that's also rapidly maturing in the open ecosystem, with projects like the LiteLLM Agent Platform offering self-hosted, Kubernetes-based infrastructure for running AI workloads with isolated sandboxes and persistent sessions.

That combination — capable open models plus mature open infrastructure — is what eventually unseats the "you must use our API" default. We saw it happen with LLMs as Llama and Mistral made it onto self-hosted stacks. Video is roughly 18 months behind that curve.

Where this is heading

Three things are worth watching over the next year if you're an Australian creator, developer, or media business trying to figure out where to place bets.

1. Quality will keep compounding in the open tier. SANA-WM at 2.6B today; expect a SANA-WM-XL or a community fork at 7B or 14B before long, closing more of the quality gap with closed models.

2. The economics will get awkward for some SaaS players. If you can self-host a model that does 80% of what a paid API does, at 10% of the long-term cost, the per-second pricing model gets harder to defend for anything but the most premium use cases.

3. Regulation and provenance will move from optional to mandatory. As open video models proliferate, the policy conversation in Australia — around deepfakes, election integrity, and synthetic media labelling — will sharpen. Open weights make watermarking harder to enforce centrally, which means provenance standards like C2PA become more important, not less.

The takeaway

SANA-WM isn't the most powerful video model in the world. It doesn't need to be. What it represents is something more durable: a credible, runnable, modifiable open baseline for AI video, at a scale a real human can actually use.

The first wave of AI video was a spectator sport — people sharing impressive demos from systems they couldn't touch. The second wave, the one SANA-WM is part of, is about handing the tools over. For creators who'd rather build than queue, that's the more interesting story.