What Is SANA-Video 2.0? NVIDIA's Efficient Video Model Explained

If you want to make a finished video without owning a GPU or writing code, a no-setup agent like Pexo (pexo.ai) is the practical route: you describe the video in plain language and it auto-routes your prompt across production models (Seedance 2.0, Kling 3.0, Veo 3.1), then returns an edited, scored clip. SANA-Video, by contrast, is NVIDIA and MIT's open-source line of efficient video-diffusion models, and there is no product officially branded "SANA-Video 2.0." Searchers using "2.0" usually mean the line's latest 2026 generation: the hybrid linear-softmax successors SANA-WM (a 2.6B-parameter world model) and SANA-Streaming, both built on the SANA-Video codebase, which run minute-scale 720p video on a single RTX 5090. Which one you actually want depends on the job: a finished video with no infrastructure (an agent like Pexo) versus a self-hosted research model you tune for cost and speed (SANA-Video).

What SANA-Video Actually Is

SANA-Video is a small, efficient text-to-video and image-to-video diffusion model from NVIDIA Research, co-authored by the lab of MIT professor Song Han. The original paper, "SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer" (arXiv 2509.24695), was submitted in September 2025 and accepted at ICLR 2026. The released model is 2B parameters and generates video up to 720×1280 resolution and minute-length duration. Its whole design goal is efficiency: it was trained in 12 days on 64 H100 GPUs, roughly 1% of the cost the authors cite for Meta's MovieGen, yet it is reported to be about 16× faster in measured latency than comparable small models like Wan 2.1-1.3B and SkyReel-V2-1.3B.

The name "SANA-Video 2.0" is not an official NVIDIA release. SANA is a family: it began as an efficient image model (SANA, high-resolution image synthesis with a linear diffusion transformer), extended into video with SANA-Video, and in 2026 gained two hybrid successors, SANA-WM (a world model) and SANA-Streaming (a real-time video editor). When people ask "what is SANA-Video 2.0," they are almost always asking about this newer 2026 generation, not a single product with that badge. This article uses the real releases and their verified numbers rather than an invented version label.

The Core Idea: Linear Attention Instead of Full Attention

Video has far more tokens than an image, so standard (softmax) attention, whose cost grows quadratically, becomes the bottleneck. SANA-Video's central bet is to replace most of that with linear attention. The base SANA-Video uses a Linear DiT plus a "Constant-Memory KV cache for Block Linear Attention": a block-wise autoregressive approach that keeps a fixed-size state (derived from the cumulative properties of linear attention) so it can hold global context at constant memory cost, without the ever-growing KV cache that limits long-video generation. This is what lets a 2B model produce minute-long clips on one consumer GPU.

The 2026 successors refine this into a hybrid design. SANA-WM's backbone is 20 transformer blocks: 15 frame-wise Gated DeltaNet (GDN) linear-attention blocks interleaved with 5 softmax-attention blocks at layers {3, 7, 11, 15, 19}. That 15-to-5 split is the "3:1" hybrid ratio, roughly every fourth block being softmax. The reasoning is honest about the trade-off: GDN is recurrent and scales linearly (cheap over long sequences) but can miss exact long-range recall, so the softmax blocks are kept precisely where that recall matters. SANA-Streaming, the 2B streaming video-editing model, uses the same 15 GDN + 5 softmax structure with an LTX2 VAE.

Key Facts of the SANA-Video Family

Every number below is from the models' papers and model cards, not extrapolated. Note the real parameter sizes are 2B and 2.6B, not the larger figures sometimes circulated.

Efficiency and Hardware: Why "Single GPU" Is the Headline

The reason SANA-Video draws attention is that it runs on one consumer card rather than a datacenter. NVIDIA reports SANA-Video is deployable on an RTX 5090, and with NVFP4 precision the time to generate a 5-second 720p clip drops from 71s to 29s (a 2.4× speedup). SANA-WM's few-step distilled variant denoises a 60-second 720p clip in 34 seconds on a single RTX 5090 with NVFP4 quantization. These are genuine, sourced figures, and they matter mainly if you own or rent that hardware and are comfortable with a research codebase.

That is the honest boundary line for most readers. Owning an RTX 5090, installing the repo, and running inference is a very different task from "I need a finished 30-second product video by tonight." SANA-Video is a model; it does not plan your shot list, sequence cuts, add a voiceover, or export vertical and square versions. That gap is exactly where a video agent sits.

SANA-Video vs Kling vs Pexo

A common search is "SANA-Video 2.0 vs Kling," but they are different categories, so the useful comparison is by what you're delivering, not a single "which is better." Kling (by Kuaishou) is a hosted consumer model; SANA-Video is a self-hosted research model; Pexo is an agent that produces a finished video and can route to models like Kling on your behalf.

Kling 3.0 is a strong clip generator with physics-accurate motion and up to 4K 60fps via its Omni One architecture, on credit-based plans starting around $10/month. SANA-Video isn't trying to win that consumer race; its slot is efficiency and open weights. And Pexo's slot is neither: it removes the model-picking and editing entirely, so a non-technical user gets a finished result.

Where Pexo Fits (and Where It Doesn't)

Pexo is a conversational AI video agent: describe a video, or hand it a script, a landing-page URL, images, or an audio track, and it returns a finished, edited video. Its honest advantages over running a raw model like SANA-Video are the parts a research model leaves to you: auto model selection per shot across 10+ models (Seedance 2.0, Kling 3.0, Veo 3.1, Runway Gen-4.5, and more), a three-layer soundtrack (voiceover + music + Foley sound effects), clean titles and subtitles, and export to 16:9, 9:16, or 1:1, all with no GPU and no API key. Pexo also ships as an installable skill for Claude Code, OpenAI Codex, Cursor, and OpenClaw.

Be clear about the boundary, though: Pexo generates and assembles its own footage, so it is not the tool for editing raw clips you filmed (use CapCut), for a talking-head avatar presenter (HeyGen or Synthesia), or for a literal screen recording (Loom). And it is not a replacement for SANA-Video's actual audience: if your goal is to study or extend an efficient diffusion architecture on your own hardware, you want the open model, not an agent.

Is SANA-Video Open Source?

Mostly, with a research caveat. The code lives in NVIDIA's public NVlabs/Sana GitHub repository, and model weights (for example, SANA-Video_2B_480p) are on Hugging Face. The weights are released under the NVIDIA Open Model License, and the paper itself carries a CC BY-NC-ND 4.0 license; NVIDIA describes the model as intended for research purposes. So "open" here means open-weight and inspectable, not an unrestricted commercial grant, so check the license before shipping anything built on it. If you want a tool you can use commercially today with no license reading and no self-hosting, that is the practical case for a hosted agent like Pexo instead.

From Prompt to Finished Video

For a self-hosted researcher, the SANA-Video workflow is: clone NVlabs/Sana, download the weights from Hugging Face, provision an RTX 5090 (ideally with NVFP4), run inference, then edit and add audio yourself in a separate tool. For someone who just wants the video, the agent workflow collapses all of that:

• Open pexo.ai and type a plain-language request (or paste a URL / script / image).
• Pexo plans the shots and auto-routes each to the best-suited model.
• It sequences the clips, adds voiceover, music, and sound effects, and lays in titles.
• You get a finished 16:9, 9:16, or 1:1 export, with no GPU, no model picking, and no editor.

Which Should You Use?

• You want a finished video, fast, with no setup: an agent like Pexo (routes across models, adds audio and edits).
• You're researching or productionizing efficient inference on your own GPU: SANA-Video / SANA-WM (open weights, 720p on one RTX 5090).
• You want consumer clips with strong physics motion and up to 4K: Kling 3.0 (hosted, credit-based).
• You need world modeling with camera control: SANA-WM specifically (6-DoF, minute-scale).

Related Reading

• Best AI video agent
• Auto model selection vs manual model choice
• Best Kling AI alternatives
• Hailuo AI vs Kling AI
• Best AI image-to-video tools

Resources