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

MiniMax H3 is the late-2026 video moment you've been hearing about: an open-weights, 33B omni-modal model that does text, image, video, and audio as one input - native stereo sound in the video, up to 15 seconds at 2K. The catch, if you're not in the licensed territory or don't have the GPU for a 42.5GB model, is that running it locally is either not allowed or not possible (closed-source-models.md covers the license geofence in detail). This node is the workaround the Zhenzhen pack offers: the H3 fast tier, served through the cheap domestic API, wrapped in a Concurrent Submit twin so you can queue it in the 10-worker video pool.

What the fast tier changes

The _fast models are the distilled/quick variants - minimax-h3-ow-t2v-fast, r2v-fast, i2v-fast - trading a little fidelity for much shorter turnaround. When you're iterating on motion or lip-sync tests, that's the right trade; save the full tier for finals.

The inputs that matter

• model - the three fast variants, text/image/reference-to-video.
• prompt - required.
• seconds - 5, 10, or 15.
• resolution - 480p or 720p (the fast tier doesn't push higher here).
• ratio - the usual set.
• image1–image9 - up to nine reference images for i2v/r2v. That's the H3 unified-context thing: throw a stack of references at it and it figures out the subject.
• audio - an audio input for audio-driven clips (the README mentions audio-driven lip-sync digital humans on this family).
• api_config - domestic route, so wire the zhenzhen-seedance2-low-price-api-settings node.
• seed / skip_error - cache-control seed and fail-silently toggle.

The concurrent flow

Output is a task (COMFLY_VIDEO_FUTURE), not a video. Stack your jobs into ComflyConcurrent_Video_Await (10 slots), which waits for all of them, returns the clips in slot order, and appends a status JSON. Because fast-tier jobs turn around quickly, the pool actually stays full - which is exactly when the concurrent design shines.

Install and gotchas

Same pack: T8mars/Comfyui-zhenzhen via ComfyUI Manager or git clone https://github.com/T8mars/Comfyui-zhenzhen, then restart. No local weights - that's the point - and standard Python deps. Paid token, overseas servers (proxy from China). One thing to keep straight: this is the reseller-hosted H3, so H3's license geofence is the author's problem, not yours - you're just buying API calls. The price is per second and per resolution, so 15s at 720p is a different line item than 5s at 480p.