Mochi-1 🍡: A Quantum Leap for Generative AI Video (Open Source!)

Mochi 1 preview is available now for Genmo users to try out for themselves

I’m really impressed by how well the high-speed camera captured the motion of the subject. The movement is incredibly dynamic and fluid!

The drawback at this point is the lower resolution, but the motion itself is just incredible.

Let us know what you think in the comments below!

Chapters 🎬🎬

0:00 - intro
0:14 - what is mochi
01:29 - who is Genmo AI
2:48 - why is Mochi 1 SOTA?
4:38 - mistral VLLM overlap
5:14 - tokenizing video
5:34 - what GPUS to RUN
5:54 - running locally
7:14 - blog
8:15 - benchmarks
9:05 - conclusion

[xbon]

the h100 part is kind of a meme. you can still run this locally with SINGLE 3090, 4090, ada 6000, etc.

[-Burs]

I resonate with open-source very well. But hearing the model needs 4x A100s, worth more than my house, my car, and my two kidneys, makes it kinda same as being closed-source :) We're getting there, yeah, but until then, it's just another project at end of my queue list. Thanks for the video tho.

[TomHimanen]

Could you create a video about finetuning generative video models? Especially Mochi would be realistic finetunable because of their appealing licence.

[fontenbleau]

every news report from Ai flux is like Christmas present 🤗

[DataJuggler]

5:43 Only $120, 000 to $160, 000 in GPU's, but yeah, its opensource.

[reinerheiner]

Regarding all those comments that complain that it only takes a lot of money to run them locally: If it is what it takes, it is what it takes. I really wonder though how much more we can optimize those algorithms to work with less VRAM and less compute. But at some point where will be a minimum required no matter what, and when that happens, the endless tears in the comment section wil be for nothing, cause it aint gonna change anything. And when that point comes, how much will the chip giants gatekeep consumer hardware so they'll be able to sell their business AI stuff for a jucy premium?

[ChristophBackhaus]

This is amazing but I think we still need improvements in STT

[cortezforever]

Will other companies use this open source model and provide it to customers?

[wonder]

I tested it on a 3090 and it is impressive. Too slow to be useful.

[thomassynths]

The first several videos you showed were brutally poor quality.