Pika put out H3 Flex, an open version built on top of MiniMax H3. Runs locally at 480p/768p with audio, and the upscaler stays on their servers

by @katyasato_art in c/hailuo-community · · Comparison

Pika H3 Flex — Faster-than-real-time video model
A post-trained H3 video model built for professional iteration: 15-second clips in as little as 7 seconds, separate audio stems, output up to 4K and open weights.
pika.art

For the Hailuo people: Pika post-trained MiniMax H3 into something they call H3 Flex, and the generator adapters plus inference code are up on Hugging Face and GitHub.

What I could pull from their page:

Local generation at 480p and 768p, with synced audio
Clips from 3 to 15 seconds, 24 fps
They claim six sampling steps instead of 50, and up to 1.8x faster than H3 Max on transformer time (their number, not end-to-end)
The 1080p / 2K / 4K upsampler is hosted only

So it's "open" in the sense that you can make drafts on your own card, and "closed" at the step where it starts looking expensive. Honestly that split might be perfect for blocking out shots before paying for finals.

Read the license before you use it for client work. I haven't tested it myself yet. Has anyone got it running?

#minimax-h3 #pika #open-weights #local

113 points · 12 comments

Comments

  • @meitanaka_fx · · 17 points

    Compared five prompts side by side with hosted H3. Flex holds composition well, but fine texture (fabric, skin pores) is noticeably softer before upscale. Motion is very close.

  • @diegoverma_art · · 6 points

    Same result here. Neon signage text gets mushy at 768p, which matters for my stuff. Fine for blocking, not for the hero shot.

  • @katyasato_art · · 6 points

    Close motion is what matters for drafts. Texture you fix in the final pass anyway.

  • @dmitrischneider_motion784 · · 12 points

    Got it running last night on a 4090. 768p, 10 seconds, about 40 seconds wall clock with the model already loaded. The first load is the slow part. Audio is there but thin, like a phone recording of the right sound.

  • @katyasato_art · · 5 points

    40 seconds for a 10 second draft is wild. That's faster than waiting in the hosted queue on a busy evening.

  • @isabellakapoor_creates · · 6 points

    On the license: the adapters sit on top of MiniMax's community license terms, and at least one catalog lists commercial use as restricted. Not legal advice, but "it's on Hugging Face" is not the same as "you can sell it".

  • @katyasato_art · · 8 points

    Yes, that's why I put the warning in the post. Drafts locally, finals through a paid route you're licensed for.

  • @jamesclarkpro105 · · 6 points

    What's the VRAM floor? Their page just says "depends on your setup", which usually means "more than you have".

  • @dmitrischneider_motion784 · · 2 points

    With offloading I saw people running 480p on 16 GB. Slow but it works. 768p at 10+ seconds really wants 24 GB.