Concurrent Submit | zhenzhen-minimax-h3-ow-video-lowprice

MiniMax H3 is the open-weights omni-modal model that finally gave video a serious open answer to Veo - native stereo audio, unified text/image/video/audio context, and genuinely strong motion. It's also a 33-billion-parameter, ~42.5GB download with a license that geofences the US, EU, UK, and South Korea out of local use. So for a lot of people, the realistic way to actually use it is exactly this: the full (non-fast) H3 tier served by the Zhenzhen pack's cheap domestic API, wrapped in a Concurrent Submit twin so you can run it in the 10-worker video pool instead of waiting on one render at a time.

The inputs that matter

• model - minimax-h3-ow-t2v, minimax-h3-ow-r2v (reference-to-video), or minimax-h3-ow-i2v (image-to-video). Same model, different input modes.
• prompt - required.
• seconds - 5, 10, or 15. The longer clips are where H3's audio sync gets genuinely impressive.
• resolution - 480p or 720p on this route.
• ratio - standard set.
• image1 - a single image reference for i2v mode. (The fast tier sibling takes up to nine images plus audio; this one keeps it simple.)
• api_config - domestic route; wire the zhenzhen-seedance2-low-price-api-settings node.
• seed / skip_error - cache-control seed, and the fail-silently toggle.

The concurrent flow

Output is a task (COMFLY_VIDEO_FUTURE), not a video. Wire your tasks into ComflyConcurrent_Video_Await (10 slots) and it waits for the whole batch, returns the videos in slot order, and appends a status JSON. If you're doing a character-consistency study - same prompt, multiple reference frames - the batch layout is natural.

Install and gotchas

Part of T8mars/Comfyui-zhenzhen: ComfyUI Manager → search "Comfyui-zhenzhen", or git clone https://github.com/T8mars/Comfyui-zhenzhen into custom_nodes, then restart. No model files to fetch - the whole point is someone else hosts them. Paid token, overseas servers (proxy from China), per-second pricing that scales with resolution. Worth remembering you're paying for API calls to a model whose weights are technically open - if you're in an allowed territory with a big GPU, the local route is the cheaper long-run option, and this node is the convenience path.