MiniMax H3

Highlights

• Video + native stereo audio generated in one pass, no separate voiceover or sound-effect model needed
• 4–15 second duration, default short side 768px, can be pushed to 2K via the Regenerate-2K workflow
• Multimodal references: ≤9 images, ≤3 videos (each 2–15 seconds, total length ≤15 seconds), audio reference
• Four modes each have official templates: T2V / I2V / reference-to-video (R2V) / first-last-frame (FLF2V)

Scope

• Good for: short films with dialogue or sound effects, product demos, talking-head material
• Good for: animating a single character image (I2V / R2V)
• Not for: projects requiring direct 4K output or hard duration requirements

Tips

• For H3 prompts, write in four parts: “shot + subject + action + sound”; the official repository has corresponding prompt structure guidance
• Reference video total length has a cap; exceeding it truncates, so cut first before feeding
• 2K is a second-pass regeneration, not native 2K; do not expect details invented from nothing

Cautions

• The community license imposes constraints on commercial use and content; the 2K workflow depends on additional components
• Generated content goes through moderation; violating prompts will be blocked

Parameters

Duration

4–15 seconds

Resolution

Default short side 768px; 2K requires Regenerate-2K

Mode

T2V, I2V, Ref2VA, FLF2V

Weight format

Official original + community pruned int8 (about 19.5GB) / 16g version (about 26.5GB)

Deployment

16GB conclusion

Must rely on offload or quantized versions: the 19.5GB int8 weights alone exceed 16GB VRAM.

VRAM note

Your machine already has minimax_h3_fl2va / ref2va int8 and 16g versions; in practice this is “runs, but slow”; keeping resolution at 768 short side and reducing frame count is most effective.

Suggested starting point

First run 5 seconds, 768p, int8 weights to get it working; for audio-video sync use the official T2V template rather than wiring nodes yourself.

Comparison

4.97 million downloads in the last 30 days, 2.4× the second-place LTX-2.5—on the video side it has no real competitor. Compared with the LTX line, it adds native stereo audio + multi-reference inputs (≤9 images + ≤3 videos), but lacks a lightweight tier that fits in 16GB: int8 weights start at 19.5GB and require offload.