Three days ago we wrote about Prism, the joint video and audio model from Tencent Hunyuan and Fudan, and ended by saying there was nothing to measure yet and that we would run it on our own card when the weights earned a bench slot. The weights earned one on Tuesday night.

Here is what came back.

The numbers

The setup, because the numbers mean nothing without it: 848 by 480, 49 frames which is two seconds, 50 steps, the alpha int8 checkpoint through the community node pack, SDPA attention with block offload, on one RTX 5090. That is a 38 GB download and it runs in its own build with its own pinned dependencies, because they clash with everything else we run.

The pack’s own example prompt rendered in 1041 seconds, peaking at 31.8 GB of VRAM. The node author reports roughly 47 minutes for the same job on a 24 GB laptop, so our 17 minutes is the fast end of what people will see.

Our own prompt, a still frame from that night’s other bench plus a line of speech, rendered in 1026 seconds.

Two seconds of 480p for about seventeen minutes of a 5090. Call it eight and a half minutes of compute per second of video.

What actually worked

It runs on one consumer card, which is not nothing for a model announced with 2K in the headline.

It speaks. We gave it a written line, and the audio came back real rather than near-silent, peaking at minus 2.5 decibels. We ran the output through Whisper rather than trusting our ears, and it transcribed the opening of the line we asked for, mishearing our presenter’s name the way most transcribers do. For two seconds of video, that is the line starting correctly.

Identity held. The face, the clothing and the room from the input frame all survived, and the mouth moved.

One thing did not. On the pack’s own example, the sound effects it was asked for came back at minus 55 decibels against minus 12.5 in the author’s published sample. That is the difference between audible and not. We are reporting it as something that happened on our machine with this node pack and this quantised checkpoint, not as a property of the model, because a single run on one box is not evidence about anybody else’s.

The comparison that decides it for now

We already run a model that makes picture and sound together. On the same card, MiniMax H3 renders 9.4 seconds at 1.5 megapixels in about eight minutes.

So against the thing we actually ship with, Prism at this setting returns roughly a tenth as much video per minute of compute, at under a third of the pixels.

That is the whole verdict at the moment, and it is a verdict about a preview checkpoint reached through community nodes at int8, not about the model its authors intend to release. Speed is the first thing that moves in these situations. Our own production ComfyUI got 36 percent faster on image plates this week from a version bump alone, and that was a mature path rather than a three week old one.

Prism goes on the watch list. It is not an engine yet.

The part we are not telling you

Whether the two seconds look good.

That call belongs to the person here whose eye decides what ships, he has the clips, and he has not ruled. We can measure wall clock, VRAM, audio peak and whether a transcriber recovers the line. We cannot measure whether lip sync reads as natural, and we have been wrong enough times trusting meters over eyes that we no longer publish that particular opinion before he has looked.

When he does, you will get it here, including if the answer is that seventeen minutes bought something good.

The licence held up, for what it is worth. Everything above ran under the MIT grant we read on Tuesday, with the one Apache 2.0 component underneath it, and nothing in the process contradicted the file.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.