Cosmos3 Nano: the NVIDIA world model for physical scenes and longer shots, and where to run it
OpenMDW 1.1, weights free, commercial use allowed. Self-hosting means a data-centre GPU: NVIDIA tested it on GB200 and H100, recommends H200, supports BF16 only, and reports a 720p clip of 189 frames in about 208 seconds on one H100. A layer-by-layer offload option exists for smaller cards. In your browser it runs on cnaps studio, included in the basic plan.
It is a world model rather than a film tool. NVIDIA trained it for physical AI, robots, vehicles and machines moving through real spaces, and besides text and image prompts it can continue a clip you already have from its last frames. It can also add a 48 kHz soundtrack when you describe the sound.
The cost is speed. At the default 189 frames a 720p clip typically takes seven to eight minutes in the studio, and sound you do not describe comes out as speech-like babble. Eight video generation models sit in the cnaps studio catalog, and this is the one for extending footage and physical scenes.
What it actually is
The 16-billion-parameter member of NVIDIA's Cosmos3 family, released in May 2026, built as a Mixture-of-Transformers around a dense 8B core. One model covers generation, reasoning and action prediction; the studio exposes the generation side. It was trained on 767 million images, 348 million videos, 139 million audio clips and 8 million action samples.
In continuation mode it anchors on the last five frames of your clip, and the output's first five frames repeat that tail, so drop them when joining the two. A video input takes priority over an image.
In cnaps studio you give it text plus an optional image or video. Ten canvases are offered across 720p and 480p, default 1280 by 720; the 480p sizes roughly halve memory and time. Length runs from 61 to 301 frames, default 189, at 16, 24 or 30 fps. Steps run from 20 to 75, default 35, and guidance from 1 to 10, default 6.
Try Cosmos3 Nano in cnaps studio →
The published numbers
NVIDIA publishes latency for the generator on single GPUs, at BF16, batch size 1, 189 frames and 24 fps. These are wall-clock seconds per clip with the reference code.
Image conditioning costs nothing extra, and dropping to 480p cuts the time by about 70%. The studio runs the same 189-frame default at 35 steps, and typically returns a clip in seven to eight minutes.
NVIDIA's technical report also scores the family against closed and open models on generation and reasoning, all measured by NVIDIA. Judge the output by eye on your own scenes.
How it compares to the other video generators here
These four share no published test, so the table maps inputs, length and sound. The deciding questions are whether you have footage to extend and whether the scene is about physical motion.
Reach for Cosmos3 Nano to extend a clip or simulate robots, vehicles and machines. Reach for LTX-2.5 for 1080p with sound, MiniMax H3 for reference-driven clips, and Wan2.2 TI2V 5B for quick, permissively licensed drafts.
What to chain it with
Text plus an optional still or clip in, a longer or new clip out, so it can sit anywhere in a video chain.
Two chains suit it. Feed a clip from Wan2.2 TI2V 5B or LTX-2.5 into the video input to carry the shot on, then join the two after dropping the repeated frames. For training data, SAM3.1 (Video) can mask the robot or vehicle in every frame of the result.
Open Cosmos3 Nano in cnaps studio, try one prompt at 480p first to judge the motion quickly, then rerun the keeper at 720p with sound switched on.
FAQ
What does Cosmos3 Nano need to run locally?
A data-centre GPU. NVIDIA tested it on GB200 and H100 and recommends H200, with BF16 as the only tested precision. A 720p clip of 189 frames takes about 208 seconds on one H100, and a layer-by-layer offload option helps smaller cards. In cnaps studio it runs in the browser, included in the basic plan.
How do I continue an existing clip?
Connect the clip to the video input and describe what happens next. The model anchors on its last five frames, ideally from 24 fps footage, and stretches it to the chosen canvas. The output starts by repeating those five frames, so trim them before joining the clips, and keep the same canvas for both.
Why does the sound come out as babble?
Sound is generated from the prompt, so the model needs to be told what to play. When the prompt describes only the picture, it fills the track with speech-like noise. Add a line such as "Sound: light rain and distant traffic", or leave sound off and add effects later with MOSS.
