What’s new
Compared with the base MiniMax H3:
- Built jointly by MiniMax and fal.ai: the official docs describe H3-Max as a model fal.ai post-trained on top of MiniMax H3, specifically optimized for generation speed.
- Faster generation: it's positioned around mainstream resolutions with higher throughput, for work where turnaround matters more than the top resolution tier.
- fal.ai says it's tuned for stronger prompt adherence and better aesthetics, while aiming not to sacrifice output quality (that framing comes from fal.ai itself).
- Resolution shifted to 480P/768P: the ceiling moved down to the more common tiers, in line with the faster positioning.
What it’s good for
- Producing a lot of material quickly — batch storyboard tests, fast creative iteration.
- Work where 480P/768P is plenty and turnaround time matters more than peak resolution.
- The same three modes as MiniMax H3 — text-to-video, image-to-video (first/last frame), and reference-to-video — just on the faster track.
Prompting tips
- Prompting works the same way as MiniMax H3: describe the relationship between your references in one sentence rather than filling in a fixed task template.
- Use bracketed camera tags like [pan] or [zoom] right after the relevant description.
- First/last-frame accepts 0, 1, or 2 images: with no image, it becomes plain text-to-video.
- Reference mode caps mixed inputs at 12 files total (images, video and audio combined) — the same per-type limits as MiniMax H3 apply.
- Duration is an integer from 5–15 seconds, a slightly higher floor than MiniMax H3's 4 seconds — don't ask for anything under 5.
Reference-to-video, 768P, 8 seconds.
The character keeps the look from reference image 1, moving to the street-dance rhythm from reference video 1.
[follow shot] keep the character centered in frame, with a neon-lit night market in the background.
Known limits
- The resolution ceiling is 768P, below MiniMax H3's 2K — use the base H3 model when top resolution matters most.
- Minimum duration is 5 seconds; MiniMax H3 itself allows as low as 4.
- The post-training approach was done by fal.ai, and MiniMax's own docs don't go into detail on the tuning method.
Using it in Nomi
On the generation canvas, add a video card, pick “MiniMax H3-Max” in the model picker and choose a mode (Text to video, First and last frame). Add your references, write the prompt and generate. In the storyboard you can pick it per shot. The first time, connect any one of the providers above under Settings → Models.
