Prism 2K video-audio needs 4+ GPUs: the hosted route up to 4K

Prism, the joint video-audio research release from Tencent Hunyuan, can render 2K (1440x2560) clips with audio, but its repository lists a single 80 GB GPU for 720p inference and four or more GPUs for 1080p and 2K. If you do not have that hardware, Sume lists a hosted id that makes native-audio clips up to 4K: Gemini Omni Flash 1.1, for 3 to 10 seconds.

Prism facts are from its GitHub repository, read 2026-10-11. Sume facts come from Video Generation and Video Router. Sume does not list Prism, and this post compares what each route needs, not output quality; I did not run either.

What the Prism page says

Prism is a dynamic sparse attention framework for training joint video-audio models natively at 2K. The repository describes a 15B-parameter model built on the MOVA base, in text-to-video-audio and image-to-video-audio modes, with structured audio prompts using tags for music, sound effects and speech. It states a 2.5x training speedup over full attention. Weights are published as preview alpha and beta checkpoints, and the page content I read did not state a license, so check the repository before any commercial use.

What Sume lists with sound up to 4K

The Video Router docs describe gemini-omni-flash-1.1 as 3 to 10 seconds at 360p, 720p, 1080p and 4K, in 16:9 or 9:16, with native synced audio that cannot be turned off. Its edit mode also accepts a source video. Sume bills the provider's per-second list x 1.25; the rate card in Sume's pricing tests is $0.03, $0.10, $0.15 and $0.30 per second at 360p, 720p, 1080p and 4K, so Sume's per-second price is $0.0375, $0.125, $0.1875 and $0.375. Those rates are from that test and can change, so confirm with GET /v1/catalog.

How the two routes differ

The gap is mostly hardware versus control.

• Length: Prism lists about 11 to 12 seconds; Omni Flash tops out at 10, so Prism can be slightly longer per take.
• Resolution: Prism reaches 2K on its own pipeline; Sume's Omni Flash goes to 4K, but that is a different model with its own look.
• Audio control: Prism's repository describes tagged audio prompts for music, effects and speech. Sume's Omni Flash has audio always on and no audio reference input, per the docs.
• Hardware: Prism needs your own 80 GB card at the lowest tier. Sume needs an API key and a balance.

When to choose which

Run Prism if you are doing research on joint video-audio generation, own multi-GPU servers, and want to examine the weights. Use the hosted id if you want a finished clip with sound from a request and do not mind that the model, its settings and its seed are not yours to change: Sume's catalog models reject a seed and any provider.options. Whichever you pick, test on your own prompt before committing a budget, because a parameter count or a resolution tier says nothing about whether a clip suits your brand.