Hailuo H3 API
Hailuo 3 视频生成模型，支持 4～15 秒文生、首尾帧图生和图片、视频、音频参考生成，输出 768p 或 2K。
What is Hailuo 3
Hailuo 3 (also MiniMax H3, Hailuo 3.0, or Hailuo 03) is MiniMax's video generation model released on 2026-07-31: a single omni-modal Transformer with official specs of 4–15s clips, 24fps, 768p/2K resolution, and native 32kHz stereo audio generated in the same pass as the visuals. MiniMax further claims up to 9 reference images + 3 reference videos + 3 reference audios (12 files total), prompts of roughly 7,000 characters, and audio in 11 languages.
Roko exposes hailuo-3 through the /v1/videos/tasks async API: text-to-video routes to minimax-h3-text-to-video, image-to-video to minimax-h3-image-to-video, and reference-to-video to minimax-h3-reference-to-video; input modalities cover text, image, video, and audio. Duration is 4–15s (default 4), resolution 768p/2K (default 768p), aspect ratios 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 (default 16:9).
Pricing per the Roko catalog: $0.0836/s at 768p and $0.1364/s at 2K, billed on output seconds; when references include video, billing covers input video seconds + output seconds (equivalent to $0.1672 per output second at 768p, $0.2728 at 2K), and each reference image from the 6th onward adds $0.0418. MiniMax's official API per-second prices differ and are not interchangeable with Roko's.
Model pricing
Pricing follows the current production configuration and the billing dimensions supported by each route.
| Route | Price | Billing rule |
|---|---|---|
| 文生 / 图生 / 图片或音频参考 | 768p $0.0836/sec2k $0.1364/sec | 按输出视频秒数；第 6 张起的参考图另计 |
| 视频参考 | 768p $0.0836/sec2k $0.1364/sec | 按输入视频秒数 + 输出视频秒数；第 6 张起的参考图另计 |
Cost estimate examples
Examples are recomputed from Roko catalog prices; estimates are made at submit time from public parameters and may be settled on the actual usage the provider reports.
8s × $0.0836/s = $0.6688
Standard no-reference tier; the common choice for batch social clips.
10s × $0.1364/s = $1.364
Upload 1–2 images as first-frame or style reference at 2K, for masters that need post-production cropping.
(6s input video + 6s output) × $0.0836/s = $1.0032
Video references bill both ways — input video seconds + output seconds — equivalent to $0.1672 per output second.
8s × $0.0836/s + 3 × $0.0418 = $0.7942
Each reference image from the 6th adds $0.0418, so 8 images means 3 surcharged; fits multi-character, multi-scene consistency.
Model information
Only provider-documented routes, capabilities, and limits implemented by this gateway are listed.
minimax-h3-text-to-video
- Limits
- —
- Duration
- 4–15s
minimax-h3-image-to-video
- Limits
- Images 1–2
- Duration
- 4–15s
minimax-h3-reference-to-video
- Limits
- Images 0–9 · Videos 0–3 · Audio 0–3
- Duration
- 4–15s
Key capabilities
Omni-modal reference input (up to 12 files)
The reference-to-video route accepts up to 9 reference images, 3 reference videos, and 3 reference audios — 12 files total — to pin characters, scenes, motion, and voice. Video and audio references are upstream claims; validate against actual task output.
Native stereo audio in one pass
MiniMax states that 32kHz stereo audio is generated alongside the visuals, no post dubbing needed, with audio support in 11 languages. Roko bills a flat per-output-second rate; audio carries no surcharge.
768p and 2K tiers
Default 768p suits batch production; 2K ($0.1364/s) fits masters destined for cropping or large screens. Clips run 4–15s (default 4), so you can match pacing to the narrative.
Six aspect ratios
21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 (default 16:9) cover cinematic widescreen through vertical short-form without forced post-generation cropping.
Long prompts and character consistency
Upstream claims prompts up to about 7,000 characters and reference-driven consistency; at launch Artificial Analysis ranked H3 #1 in video editing, #2 in text-to-video, and #3 in image-to-video (third-party ranking, not Roko-measured).
Main parameters
Parameters come from structured model configuration; unsupported or unconfigured fields are omitted.
modelstringSubmodel ID
Default:minimax-h3-text-to-video / minimax-h3-image-to-video / minimax-h3-reference-to-videopromptstringHailuo 3 视频生成模型，支持 4～15 秒文生、首尾帧图生和图片、视频、音频参考生成，输出 768p 或 2K。
Default:RequireddurationintegerOutput video seconds
Default:4–15squalityenumCurrent billing
Default:768p / 2kaspect_ratioenumCapability
Default:21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16image_urlsarrayImage to video: 1–2 · Reference to video: 0–9
Default:Varies by routevideo_urlsarrayReference to video: 0–3
Default:Varies by routeaudio_urlsarrayReference to video: 0–3
Default:Varies by routeCapabilities and usage guidance
Key features
- Text to video · minimax-h3-text-to-video
- Image to video · minimax-h3-image-to-video
- Reference to video · minimax-h3-reference-to-video
- 768p / 2k output
- 4–15s Duration
Limits
- Image to video: Images 1–2
- Reference to video: Images 0–9 · Videos 0–3 · Audio 0–3
- Capability source: https://evolink.ai/zh/hailuo-3
Limits and caveats
- References containing video bill both ways — input video seconds plus output seconds — so longer reference clips raise cost directly.
- Reference images from the 6th onward add $0.0418 each; large image sets scale cost linearly.
- A successful submission only means the task was accepted, not that generation will succeed; the terminal state is authoritative.
- Output links expire after a limited window — download and re-host your footage promptly.
- 2K generation typically takes longer than 768p; leave scheduling headroom for time-sensitive jobs (no vendor-published timing).
How to call it
- POST /v1/videos/tasks with model set to hailuo-3, plus prompt, first-frame image, reference videos/audios, duration, resolution, aspect_ratio, and route-specific parameters; the response returns a task_id.
- Poll GET /v1/tasks/{task_id} until the status reaches a succeeded or failed terminal state.
- On succeeded, read the video URL from the task result; Roko output links expire, so re-host the footage in your own storage promptly.
- On failed, read the error message, fix the prompt or parameters, and resubmit through the same flow; failed tasks are settled per the platform refund rules.
- A successful submission only means the task was accepted, not that generation succeeded; the terminal state is authoritative.
Typical use cases
Social clips with dialogue and SFX
Input: a scene prompt (MiniMax claims audio support in 11 languages) + 16:9 or 9:16; action: text route, 4–15s; expected: the upstream capability can generate synchronized picture and sound in one pass; limits: Roko's catalog exposes no separate audio toggle, so verify the task output, and each clip is capped at 15s.
Multi-reference character shorts
Input: character reference images + scene reference video + voice reference audio (≤12 files total); action: reference-to-video route; expected: consistent look and voice across shots; limits: consistency depends on reference quality — verify empirically.
Video editing and restyling
Input: an existing clip + an edit-instruction prompt; action: video-reference route, billed on input video seconds + output seconds; expected: upstream ranking placed H3 #1 in video editing at launch (third-party, not Roko-measured); limits: input video seconds count toward cost, keep reference clips short.
2K ad masters
Input: 1–2 product images + a high-detail prompt; action: image route + 2K tier ($0.1364/s); expected: a high-resolution master ready for grading and cropping; limits: 2K costs more per second and typically takes longer — validate prompts at 768p first.
Choosing against other video models on Roko
Positioning and billing comparison only; final decisions belong to each model's catalog and your own testing.
hailuo-3 vs kling-o3
H3 leans on omni-modal reference input (images/video/audio in one shot) and native stereo audio, priced per output second; kling-o3 leans on 4K and multi-shot narrative. Pick H3 for reference-driven, audio-inclusive shots and kling-o3 for high-resolution long-form — the pricing bases differ.
hailuo-3 vs kling-3.0-turbo
kling-3.0-turbo trades on speed and a lower unit price for fast drafts; H3 offers deeper reference modality (up to 12 files) with native audio. Drafting on turbo and finishing key shots on H3 is a common split.
hailuo-3 vs happy-horse-1.1
happy-horse-1.1 favors multi-shot narrative at a lower unit price; H3's differentiation is reference depth and same-pass audio. Choose happy-horse-1.1 for budget-sensitive batch narrative and hailuo-3 for reference-driven, sound-picture-unified single shots.
FAQ
What model ID and routes does hailuo-3 use on Roko?
Submit with model set to hailuo-3: text-to-video routes to minimax-h3-text-to-video, image-to-video to minimax-h3-image-to-video, and reference-to-video to minimax-h3-reference-to-video. The actual submodel is whatever the task response reports.
How much does one Hailuo 3 video cost?
Billing is per output second: $0.0836/s at 768p and $0.1364/s at 2K; references containing video bill input video seconds + output seconds, and each reference image from the 6th adds $0.0418. For example, an 8s 768p text-to-video clip is $0.6688. Estimates use public parameters at submit time and may be settled on actual usage.
Which durations, resolutions, and ratios are supported?
Duration 4–15s (default 4); resolution 768p/2K (default 768p); ratios 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 (default 16:9). The current Roko catalog is authoritative.
Can it generate video with audio?
MiniMax states that H3 can generate 32kHz stereo audio in the same pass as the visuals, with support for 11 languages. Roko's current catalog exposes no separate audio toggle, so verify the actual task output; its pricing expression also lists no audio surcharge.
How many reference files can I send?
Image-to-video takes 1–2 images; reference-to-video takes up to 9 reference images, 3 reference videos, and 3 reference audios — 12 files total. Reference images from the 6th onward add $0.0418 each.
Why can't I get the video right after submitting?
hailuo-3 runs on the /v1/videos/tasks async API: a successful submit only means acceptance, and generation needs queue and inference time. Poll GET /v1/tasks/{task_id} and read the video URL once the status reaches succeeded.
Is there a free tier?
hailuo-3 is pay-as-you-go and the Roko catalog lists no free quota; cost is estimated from public parameters at submit and settled on actual usage, subject to the terminal task state and the pricing page.
Can I use the output commercially?
Commercial use depends on your rights to the input material, the model license, and the target platform's rules: the upstream H3-Base weights ship under a Community License — review the license and platform terms yourself before commercial use. Roko does not provide legal advice.
Do output links stay valid forever?
No. Roko output links expire after a limited window; download and re-host the footage to your own storage as soon as the task completes.
Sources
Upstream capability claims come from MiniMax's official release and open-weight materials; Roko-callable parameters, routes, and prices come from the current production catalog. Verification date: 2026-10-09.
- MiniMax Research: MiniMax H3 launch Official launch: multimodal context, up to 15 seconds at 2K, native stereo audio, and open-weight plan
- MiniMax API Platform: H3 pricing Official API rate card: 768p/2K output seconds, input-video seconds, and reference-image surcharge
- ComfyUI Wiki: MiniMax H3 open weights Open-weight release: H3-Base 768p, 24fps, aspect ratios 21:9 to 9:16, 32kHz stereo audio
- Roko API video task docs Async submit, task query, and billing lifecycle on Roko
- Roko API model catalog Current routes, prices, and generation config
Models in this family
Explore other available variants from the same model family.
Hailuo H3
Other video models
Explore more models available through the unified asynchronous video task API.
