# MiniMax H3: what the free video model actually does, and what 8 GB really buys you
China's MiniMax put the weights for its H3 video model on Hugging Face on 3 August 2026, three
days after announcing it as a paid API. You can pull them right now, without a gate, from
huggingface.co/MiniMaxAI/MiniMax-H3. The code sits
at github.com/MiniMax-AI/MiniMax-H3.
That much is simple. Most of what gets said about the model after that is not.
What it is
H3 is an omni-modal generator. It reads text, images, video and audio in one request and returns
5 to 15 seconds of video at up to 2K with native stereo sound — speech, lip sync and music
included, in a single pass rather than a video model bolted to a separate audio model. You can
pass up to nine images, three video clips and three audio tracks together.
Two variants ship. H3-Base-FL2VA handles first-and-last-frame work, which covers plain
text-to-video as the zero-image case. H3-Base-Ref2VA is the reference mode, and it is the one
that takes video and audio as input.
The 8 GB claim, straightened out
You will hear that you can run it on 8 GB of VRAM. That is true, and it is also not MiniMax's
doing. MiniMax's own repository recommends four GPUs for SGLang inference. The consumer path is
WanGP, a community project aimed squarely at people
without datacentre hardware. On v12.41 it runs a 5-second clip at 832x480 in 5-6 GB, and a
15-second clip in 8-9 GB. It supports cards as old as a GTX 10-series.
There is also a Pinokio one-click launcher that wraps ComfyUI and a pruned, quantized 63 GB build.
It is easier, and it asks for 32 GB of VRAM to be comfortable. If you have 8 GB, WanGP is the one
you want. Either way, 8 GB buys you 832x480 and softer detail, not 2K.
Two claims worth checking before you repeat them
"It beats Google's Omni model on the leaderboards." Depends entirely on which board. On
Artificial Analysis as of 31 August, Google's Gemini Omni Flash sits second in text-to-video at
an Elo of 1,237, ahead of MiniMax H3 at 1,227. H3 leads the Video Edit Arena outright at 1,390,
32 points clear, and edges Omni Flash in image-to-video-with-audio by six points. So: first in
editing, competitive in image-to-video, behind in text-to-video. Worth adding that Wan 3.0 — also
open — tops the text-to-video board at 1,242, above both of them.
"It's uncensored." The model card says otherwise. It documents safety guardrails, automated
moderation of submitted text, images and video, and industry-standard filtering, with content
suspected of being unlawful or pornographic subject to blocking. No official MiniMax source
describes the model as uncensored.
The licence word people get wrong
H3 is open-weight, not open-source. It ships under the MiniMax H3 Community License, a custom
agreement rather than an OSI-approved one, and commercial use runs through an application form
scoped to the USA, EU, UK and South Korea. The weights are free to download. That is not the same
as the licence being free of conditions, and if you are building a product on it, the difference
is the part that matters.
Worth your weekend?
If you make video and own a reasonable GPU, yes. A model that emits synced audio in the same pass
as the picture removes a genuinely annoying step, and the fact that a community build has it
running in 8 GB three weeks after release says something about how fast the optimisation layer
moves now. Just go in knowing it is open-weight rather than open-source, moderated rather than
unfiltered, and that the leaderboard story is more interesting than the headline version.