MiniMax H3 Max Turbo

The other H3 tier

Same model family as MiniMax H3 Fast, one notch up: MiniMax H3 Max Turbo. H3 is MiniMax's 33B omni-modal video model - native stereo audio, not an audio pass bolted onto a silent clip - and it's the one that took #1 in Video Editing (With Audio) on Artificial Analysis when it opened its weights in August 2026. The open weights come with a catch the KB's panel spells out: the community licence excludes the US, EU, UK and South Korea, so for a lot of readers the hosted API is the only way to touch this model at all. Locally it's ~42.5GB before you've loaded anything else.

What you're buying here over the Fast tier is resolution and a different generation recipe: this node offers 768p, and it exposes a prompt-expansion dial the Fast node doesn't.

What goes in

positivePrompt is required. Everything else is optional.

frameImages is an IMAGE socket - the image-to-video path, and the only image input this tier has. Note what's not here compared to H3 Fast: no referenceImages socket, no referenceVideos socket, no advanced_json. So Max Turbo is a straighter generator: text-to-video, or first-frame-driven video, and that's the whole input surface.

size is where the resolution lives: fifteen options, from None (model default) through 480p and 768p at six aspect ratios each, plus 480p (from input) and 768p (from input) for when a frame image should dictate the shape. Default is 480p (~16:9). Pick a (from input) entry and the frame image owns the aspect ratio - good for keeping a source photo's composition intact.

duration is 5–15 seconds (note the 5-second floor, higher than Fast's 4). Frames come from duration × the model's frame rate, per the tooltip. seed and numberResults (up to 4) are the usual pair.

Then the dial that deserves actual thought: settings.promptExpansion - disabled, balanced (default) or quality. Its tooltip ends with the line that matters: "Affects reproducibility." A model that rewrites your prompt before generating will not give you the same clip twice from the same seed, even if it usually gives you a better-looking one. Turn it to disabled while you're dialling in a composition; put it back to balanced when you finally like the shot and want it prettier.

The rest is housekeeping: the safety group (safety.checkContent, safety.mode fast/full), the ttl toggle that gates ttl_value, outputFormat (MP4/WEBM/MOV) and outputQuality (20–99).

Output: one video.

Install

Search Runware in ComfyUI Manager - it's the pack's display name, so it comes up on the first try - or do it by hand:

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

Dependencies are runware-sdk>=1.6.5, pillow and soundfile. No weights, no model folder, no VRAM ceiling to worry about - inference happens on Runware's hardware. Key from runware.ai/api-keys into ComfyUI Settings → Runware API key, or RUNWARE_API_KEY, or runware auth login. The title bar reports each run's cost when it finishes.

Where people get burned

Turning promptExpansion on "to help" and then chasing a seed. You'll never converge: the prompt you queued isn't the prompt that ran. This is the most common self-inflicted wound on any model with automatic prompt rewriting, and it's a one-click fix.

Expecting the Fast tier's reference-video inputs. They aren't on this node. If you want @tagged video references, that's H3 Fast; if you want the pixels, it's here. Neither node gives you both.

A file path where you expected a VIDEO link. The pack only produces a native VIDEO output when your ComfyUI has the comfy_api video class available - otherwise you get the saved path, with the mp4 still in your output folder as runware_<id>.mp4. Update ComfyUI, restart, and the socket comes back.

Blaming the node for a refusal. Moderation is the model's, not the pack's. There's a safety.mode and a safety.checkContent, and behind them, MiniMax's own policy - the same wall any hosted model puts up, and not one you can tune away.