MiniMax H3 Video Spicy API

MiniMax H3 (Hailuo 3.0) is MiniMax's general-purpose omni-modal audio-video generation model, released on July 31, 2026, and built on a single-stream H3-Omni-Transformer of roughly 33B parameters that jointly predicts video and audio latents — sound effects, ambience, and multilingual character lip-sync are natively generated in the same pass (32kHz stereo) rather than dubbed afterward. Each run generates 4–15 seconds at a fixed 24fps with direct 2K cinematic output, supporting text-to-video, first/last-frame control, and multimodal reference-based generation. Model weights were opened on August 2, 2026 under the MiniMax H3 Community License.

The model is offered through the iCreat platform as an API using a two-step asynchronous task flow: submit a task to obtain a task_id, then poll the result endpoint until a terminal status. It supports 768P and 2K resolutions, billed by output video duration: $0.10 per second at 768P and $0.14 at 2K; aspect ratios include adaptive plus six concrete values.

Model Positioning

MiniMax H3 is positioned as the complete multimodal audio-video generation system. It accepts mixed input of text, images, video, and audio within one unified context, then jointly generates video with native stereo sound — the full-capability form of the MiniMax H3 family: fast-iteration needs go to the post-trained H3 Max (768P only), while HD delivery, multimodal referencing, and integrated audio-video belong to the standard H3. The official reference-media ceiling is up to 9 images, 3 videos, and 3 audio clips, suited to character-consistency content, reference-driven generation, and full-narrative work like commercial advertising and short dramas.

Core Capabilities

Omni-Modal Understanding and Joint Audio-Video Generation

Text, images, video, and audio form one unified context in any order; the model jointly predicts video and audio latents, generating dialogue, ambience, sound effects, and picture synchronously — audio-video sync is built into the model, not aligned in post.

Direct 2K Cinematic Output

Rather than relying on conventional post-hoc super-resolution, the model improves resolution through self-generation, better preserving structure and text detail and mitigating the text-corruption issues common in AI video.

First/Last-Frame Control

Provide first- and last-frame images to anchor the opening and closing shots, generating controlled transitions — suited to storyboard continuity and cut design.

Multimodal Reference-Based Generation

Reference images, videos, and audio can all drive generation (annotated via role, e.g. reference_image); the official ceiling is up to 9 images + 3 videos + 3 audio clips — suited to character-consistency content and asset reuse.

Motion Fluidity and Camera Control

Exceptional motion smoothness, realistic physical simulation, and camera-work scheduling, with improved complex-instruction following over the previous generation.

Pricing

Total Cost = Unit Price × Output Video Duration. Only the output video duration is billed. The costUSD field in the response returns the actual cost once the task succeeds.

Application Scenarios

• Commercial advertising: audio-video-integrated brand films and product motion showcases
• Short-drama creation: multi-shot narrative with lip-synced dialogue
• Social media video: vertical short clips and topical material at scale
• Character-consistency content: IP characters appearing across scenes driven by reference media
• First/last-frame transitions: storyboard continuity and controlled cuts

Model Comparison

Same-Series Comparison

Cross-Model Comparison

Same-series data comes from MiniMax's official releases and channel documentation; this model's specifications follow the iCreat channel documentation.

Why Choose MiniMax H3?

• Joint audio-video generation is built into the architecture: dialogue, ambience, and effects come out in the same pass — no dubbing post-production
• Direct 2K output with self-generation-based enhancement, preserving text and structural detail better
• Unified understanding of mixed text, image, video, and audio input; multimodal reference-driven generation in one step
• First/last-frame control makes storyboard continuity and transitions predictable
• Weights are open (MiniMax H3 Community License) — cloud API and self-hosting routes run in parallel

Specifications

Architecture

The core of MiniMax H3 is a single-stream H3-Omni-Transformer of roughly 33B parameters that encodes text, vision, and audio into a unified multimodal sequence and jointly predicts video and audio latents. On the encoding side, H3-Encoder reuses Qwen3-VL-32B pretrained weights; vision passes through the temporally causal H3-VisualVAE (16× spatial, 4× temporal compression), and audio through H3-AudioVAE, which processes left and right channels before recombining them into stereo, with 3D Multimodal RoPE for positioning. The complete system consists of H3-Context-IR (multimodal instruction parsing), H3-Base (core 768P audio-video generation), and H3-Regenerate-2K (high-resolution regeneration). The model outputs fixed 24fps video with 32kHz native stereo audio; weights were opened on August 2, 2026 (dual-checkpoint repository of about 354GB, MiniMax H3 Community License).

Notes

• Submit and query must be chained with the same task_id
• While processing, result is []; keep polling, with a suggested interval of 2–5 seconds
• FAILED is terminal; check request parameters and reference media before resubmitting
• content must contain at least one non-empty text item
• Image, video, and audio URLs must be HTTPS and accessible by the server
• The role annotations of multimodal items should match their intended use in the prompt
• ratio defaults to adaptive; text-to-video requests are advised to specify a concrete ratio
• Quantity and duration constraints on reference media follow the official documentation
• costUSD is returned only when the task succeeds; failed tasks do not produce a cost field

FAQ

How do I get started with the MiniMax H3 API?​

Register on the iCreat platform and obtain an API Key from the console, send a generation request to the submit endpoint, then poll the result endpoint with the returned task_id. All requests are authenticated with the Authorization: Bearer header. Keep your API Key safe and never expose it in client code or public repositories.

How is the cost calculated?​

Total Cost = Unit Price × Output Video Duration. 768P is $0.10 per second and 2K is $0.14 per second. For example, a 5-second 2K video costs $0.70. The costUSD field in the response gives the actual cost once the task succeeds.

What generation modes are supported?​

Text-to-video (content with text only), first/last-frame control (first- and last-frame images anchoring the opening and closing shots), and multimodal reference-based generation (adding image, video, or audio reference items to content with role annotations) — all three modes share one endpoint.

How do I choose an aspect ratio?​

ratio supports adaptive plus six concrete values: 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, defaulting to adaptive. For text-to-video, explicitly specifying a concrete ratio is recommended for more predictable framing.

How do I query the task result?​

Send a POST request to https://api.icreat.ai/v1/task/result with the task_id in the body. A status of SUBMITTED means processing, with result as [] — keep polling. When status is SUCCEEDED, result returns the video resource array; read url or download_url to retrieve the video.

What if the task fails?​

FAILED is a terminal status. Check in order: whether content contains at least one non-empty text item, whether reference media URLs are HTTPS and accessible, whether resolution is 768P or 2K, and whether duration is within 4–15. Fix any issue and resubmit. The cost field is returned only when the task succeeds.