MiniMax H3 Fast

Why this node exists, in one paragraph

MiniMax H3 is the 33B omni-modal video model that opened its weights in August 2026 - the first credible open-weight answer to the audio gap in video generation, since it generates stereo sound natively instead of bolting an audio pass onto a silent clip. It also landed at #1 in Video Editing (With Audio) on Artificial Analysis. Two things follow from that. First, if you're in the US, EU, UK or South Korea, the downloadable weights are not licensed to you - the community licence geofences those regions - so the hosted API is the legitimate route to the same model. Second, even where you are licensed, 33B of omni-modal video is not a casual local load. H3 Fast is the small, cheap tier: 480p, and it's the sane way to test whether H3's motion-plus-audio is worth your money before you commit to anything heavier.

What goes in

positivePrompt is the only required input, which makes it one of the friendlier video nodes in the pack.

The interesting optional inputs are the reference sockets, because H3 is built to take context rather than just a prompt:

• referenceImages and frameImages are IMAGE sockets. Frame images are the image-to-video path; reference images are context to keep something consistent.
• referenceVideos is a RUNWARE_REFERENCEVIDEOS socket, fed by the Runware Reference Videos builder in the Runware/Params menu. That builder takes a video (a UUID or URL - max 15 seconds, and it wants a real minimum pixel count), a type of reference or extend, and a tag that must start with @ such as @video1. You then mention that tag in your prompt. It's an elegant mechanism once you see it, and completely opaque until you do.

Then the shape of the output: size offers only 480p options (plus "None (model default)"), defaulting to 480p (1:1). duration runs 4–15 seconds and the tooltip is explicit that frames = duration × the model's frame rate. seed is the usual uint32, numberResults gives you up to 4 variations per call.

Two escape hatches worth knowing. The safety group (safety.checkContent, safety.mode fast/full) turns on moderation, which per its tooltip increases total generation time. And advanced_json is the pack's raw-JSON merge - its tooltip names exactly what it's for here: inputs.referenceAudios. If H3 can take reference audio for this endpoint, that's the door, and the merging means your JSON wins on conflict.

Output is one video, in ComfyUI's native VIDEO type.

Install and keys

Manager route: search Runware, install, restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

That's runware-sdk>=1.6.5, pillow, soundfile - the pack contains no model weights, so there's nothing to download and nothing for your GPU to do. Then either paste your key into ComfyUI Settings → Runware API key, export RUNWARE_API_KEY, or run runware auth login once. Cost per run shows up on the node's title bar afterwards.

Where people get burned

You can't get 768p out of this node. That's not a bug - it's the tier. If you need more pixels, that's MiniMax H3 Max Turbo, same family, more money.

Duration is the bill. Video pricing scales with what you generate, so a 15-second 480p clip is a very different receipt from a 4-second one. Test at 4 seconds, then extend the prompt to the length you actually want.

The output may land as a file path, not a VIDEO link. The pack wraps the downloaded result as a native VIDEO only when your ComfyUI exposes the comfy_api video class; older builds hand back the path string. The clip is still sitting in your output folder as runware_<id>.mp4 - update ComfyUI, restart, re-queue.

"No Runware API key" means nothing is set in any of the three places. Check the Settings panel first; the env var overrides it, which occasionally confuses people who set both.