MiniMax H3 takes images, video clips and audio tracks in one request — point at them in the prompt by order, as Image 1, Video 1, Audio 1.
MiniMax H3 is the video model MiniMax ships inside its Hailuo 3 product, integrated here on 12 August 2026. Three endpoints sit behind one picker — text-to-video, image-to-video and reference-to-video — sharing the same panel and the same per-second price. Reference mode is the part worth learning: up to nine images, three video clips and three audio tracks in a single request, pointed at in the prompt by order rather than with @-tags.
Now the parts that cost you something. Reference video is billed: its input seconds are added to your output seconds, so a 10-second reference on a 10-second clip bills as twenty. Image-to-video uses the first frame only — the last-frame input the upstream API exposes is not enabled here. There is no 1080p and no 4K, audio has no off switch, submitted tasks cannot be cancelled, and we have not timed this model.
MiniMax's flagship video model generates 4–15 second clips with native audio at 768p or 2K, driven by text, a first-frame image, or mixed image, video and audio references — no API key, no separate account.
Text-to-video for pure prompting, image-to-video to animate a first frame, and reference-to-video for the ambitious briefs: up to 9 images, 3 video clips and 3 audio tracks in one request, referenced in the prompt as "Image 1", "Video 1", "Audio 1". Replace a person in a clip, keep a face from a photo, or drive dialogue from a voice track.
H3 outputs video with native audio, and in reference mode it can perform a supplied voice track with synchronized lip movement — hand it a portrait and an audio clip and describe the delivery. Reference audio is free; only video references add billable input seconds.
Two output tiers with per-second pricing shown before you generate: 55 credits per second at 768p, 90 at 2K. A 4-second 768p draft is 220 credits; when the motion works, rerun the same brief at 2K for delivery. Six aspect ratios from 9:16 vertical to 21:9 ultrawide.
H3 accepts prompts up to 7,000 characters, enough to describe camera movement, blocking, wardrobe and mood in one pass. Every generation keeps its full prompt and parameters in history, so a shot that works is reproducible next week.
Generated here on 15 August 2026, first take, exact prompts and settings shown. Each clip targets one of the capabilities claimed further down this page.
A thunderstorm rolling over a wheat field at dusk, wind bending the crop in waves, distant lightning illuminating the clouds, low rumbling thunder, slow push-in.
A street musician playing an upright bass on a rainy night pavement, fingers moving on the strings, reflected neon in the puddles, handheld documentary camera.
It is the reference-heavy model in this lineup: more input assets per request than the Veo tiers or Kling accept, and clip lengths past the eight-second wall.
Nine images, three video clips and three audio tracks in one request, cited by order in the prompt. That is enough to hold a face from one photo, a location from another and a voice from a recording, in a single pass.
Both Veo tiers stop at eight. H3 runs to fifteen, any whole second, at a flat per-second rate — so length is a straightforward decision rather than a jump between packages.
In reference mode it performs a voice track you supply with synchronized lip movement, and the reference audio itself is free. Describe the delivery in the prompt the way you would direct a read.
Two tiers, one flat rate each, same prompt. Run the motion cheaply until it works, then rerun the identical brief at 2K for the version you hand over.
Audio has no off switch, the ladder stops at 2K, and a submitted task runs to completion. If you need a silent clip, Kling 3.0 leaves sound off by default; if you need 4K, both Veo tiers and Kling 3.0 have that tier.
Same workspace, same credit pool — switching between any of these is one click.
| MiniMax H3 | Veo 3.1 Pro | Veo 3.1 Fast | Kling 3.0 | |
|---|---|---|---|---|
| Resolutions | 768p · 2K | 720p · 1080p · 4K | 720p · 1080p · 4K | 720p · 1080p · 4K |
| Clip length | 4–15s, any whole second | 4, 6 or 8s | 4, 6 or 8s | 3–15s, any whole second |
| Sound | Native and always on — no toggle | Optional, and it changes the rate | Optional, and it changes the rate | Optional toggle, off by default |
| Credits per second, 768p and 2K | 55 at 768p · 90 at 2K | Not offered — this tier runs 720p, 1080p and 4K | Not offered — this tier runs 720p, 1080p and 4K | Not offered — this model runs 720p, 1080p and 4K |
| Credits per second, 720p | Not offered | 80 silent · 160 with audio | 50 silent · 75 with audio | 45 silent · 60 with sound |
| Credits per second, 1080p | Not offered | 80 silent · 160 with audio — the same rate as 720p | 50 silent · 75 with audio — the same rate as 720p | 60 silent · 85 with sound |
| Credits per second, 4K | Not offered | 160 silent · 240 with audio | 150 silent · 175 with audio | 200, sound included at no extra cost |
| Credits for one clip, floor to ceiling | 220–1,350 for output alone; 2,820 with 15s of reference video and 9 reference images | 320–1,920 | 200–1,400 | 135–3,000 |
| Reference assets per request | Up to 9 images, 3 videos and 3 audio tracks, 12 in total | One image | One image | One image |
| Speed we measured ourselves | 3m 33s for 5s and 7m 32s for 20s, both 720p (2 clips, August 2026) | Not measured | Not measured | Not measured |
Resolutions, clip lengths and every credit figure in this table are read from our own model registry on 20 August 2026 — the same per-second rate tables that price your generations, not vendor marketing pages. Rates are per second of output, so multiply by the clip length you pick; on H3, reference-video input seconds are added to that count at the same rate. The resolution rows are split because H3 offers 768p and 2K while the other three offer 720p, 1080p and 4K — a tier a model does not sell is marked "Not offered" rather than filled with a nearby number. Cells we have not measured say so.
Compare MiniMax H3 with other video models — same workspace, one click to switch.
| Model | minimax-h3 (MiniMax H3, the model behind Hailuo 3) — integrated here on 12 August 2026 |
|---|---|
| Modes | Text-to-video, image-to-video from a first frame, and reference-to-video from a mixed asset set — three endpoints behind one picker, priced identically |
| Resolution | 768p and 2K — there is no 1080p tier and no 4K tier |
| Duration | 4 to 15 seconds, any whole second (4 by default) |
| Aspect ratios | 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9 in text and reference modes; image-to-video follows the reference frame |
| Reference assets | Up to 9 images, 3 video clips and 3 audio tracks, 12 in total. Reference video runs 2–15 seconds per clip and 15 seconds combined. Cited in the prompt as "Image 1", "Video 1", "Audio 1" |
| Prompt length | Up to 7,000 characters |
| Audio | Native and always on — there is no audio toggle |
| Credits | Per second: 55 at 768p, 90 at 2K, the same rate in all three modes. Reference-video input seconds bill at that same per-second rate; reference images past the fifth add 30 each; reference audio is free |
| Credit range per clip | 220 for a 4-second 768p clip, 1,350 for a 15-second 2K clip. Loaded up with 15 seconds of reference video and 9 reference images, a 2K request tops out at 2,820 |
| Cancellation | Submitted tasks cannot be cancelled |
| Measured speed | Not measured — we have not run a timed benchmark on this model |
Three steps from a blank prompt box to a finished clip.
Open the model picker and choose MiniMax H3. The panel shows its credit cost, resolution options, and duration limits before you spend anything.
Write the subject, the action, and how the camera behaves. Attach a first frame if you want the clip to start from a specific image.
Download the clip, extend it, or send it through further editing. Prompt, model, and settings stay attached in your history.
One account, the full image and video generation workflow.
Run MiniMax H3 on the free tier's daily credits, or upgrade for premium models, faster generation, and 4K output.
Perfect for getting started
Billed annually · $240/year
Billed annually · $380/year
Cinematic character performance and multimodal reference work — animating a first frame, keeping identities from reference photos, and performing supplied audio with synchronized lips. Output is 4–15 seconds at 768p or 2K with native audio.
55 credits per output second at 768p and 90 at 2K, so a 4-second 768p clip is 220 credits. Reference-video input seconds bill at the same rate; reference images beyond 5 add 30 credits each; reference audio is free. The exact total is shown before you generate.
Yes — H3 is the model MiniMax ships in its Hailuo video product. On CreateVision AI you run it directly in the browser with no MiniMax account or API key.
Up to 9 images, 3 video clips (2–15 seconds each, 15 seconds combined) and 3 audio tracks, at most 12 assets in total. Refer to them in the prompt by order: "Image 1", "Video 1", "Audio 1".
Whole seconds from 4 to 15, at 768p or 2K. Text and reference modes offer 16:9, 21:9, 4:3, 1:1, 3:4 and 9:16 (reference mode can also auto-adapt); image-to-video follows the ratio of your first frame.
No. H3 generates native audio on every run and there is no toggle for it, so the rate you see is the only rate. If you need a silent plate you can either mute it in your edit, or pick a model whose sound is optional — Kling 3.0 leaves it off by default.
H3 is the stronger value for 4–15 second character and dialogue shots. Seedance 2.5 goes up to 30 seconds and accepts far more reference assets per request, at a higher per-second price. For cheap drafts, Seedance 2.0 Mini remains the budget pick.
One-click tools for everything around generation — polish, restyle, and share.
