Wan 3.0 or MiniMax H3 Max: which to pin for reference-to-video
Both take image, video and audio references on Sume. Wan runs 2 to 30 seconds with 5 reference videos; H3 Max runs 5 to 15 seconds with always-on stereo audio.
Pin Wan 3.0 when you need a clip longer than 15 seconds, a 2-second floor or a silent track; pin MiniMax H3 Max when you want native stereo audio and 1080p in a 5 to 15 second clip. Both accept reference images, videos and audio on Sume. This is a comparison of what the catalog publishes, not of how the output looks, which you have to judge on your own prompts.
Side by side
List prices are from the catalog's dated notes (Wan 2026-08-24, H3 Max 2026-09-08) and exclude Sume's multiplier. H3 Max's note excludes a temporary launch promotion.
| Spec | wan-3.0 | minimax-h3-max |
|---|---|---|
| Duration | 2 to 30 s | 5 to 15 s |
| Resolutions | 480p, 720p, 1080p | 480p, 768p, 1080p (latent refinement from 768p) |
| Reference images | up to 10 | up to 9 |
| Reference videos | up to 5, 15 s total | up to 3, 2 to 15 s each, 15 s combined |
| Reference audio | up to 5, 15 s total | up to 3, 2 to 15 s each, 15 s combined |
| Audio | optional | always on, stereo |
| Aspect ratios | no 21:9 | includes 21:9 |
| List price | $0.05, $0.10, $0.20 per s at 480p, 720p, 1080p | $0.05, $0.08, $0.16 per s at 480p, 768p, 1080p |
How to decide
If neither is clearly better for a shot, run the same prompt on both at 480p, which is the cheapest rung for each.
- Need more than 15 seconds: Wan.
- Need 21:9: H3 Max.
- Need your own soundtrack in post: Wan with generate_audio off, since H3 Max always makes audio.
- Many reference clips: Wan takes 5 videos, H3 Max 3.
- Inspect 1080p image-to-video openings on H3 Max, as the catalog warns some outputs have a crop jump after two frames.
Sources
Related posts
More in Comparisons
- Wan 3.0 or Seedance 2.5: which is cheaper for a 10-second 720p clip?
Wan 3.0 is cheaper: a 10-second 720p clip is $1.00 at fal list ($1.25 on Sume) against about $4.62 ($5.78 on Sume) for Seedance 2.5, roughly 4.6x.
- Wan 3.0 vs Gemini Omni Flash at 720p: both list $0.10 per second
Wan 3.0 (fal) and Gemini Omni Flash (Google, about $0.10 per second at 720p) share a list price; Sume bills $0.125 per second for either.
- Which image and video models on Sume have open weights?
Sume's catalog is hosted. Where vendors publish weights for families Sume lists (Qwen-Image, MiniMax H3, FLUX.2) and why a null hugging_face_id proves nothing.
- Which image model follows a rough sketch best? A test plan on Sume
No honest ranking exists without your sketches. Run one sketch and one prompt through four Sume image models, score six traits, and read the cost per try.
Written by Sume