MiniMax H3 Launches as a General-Purpose Multimodal Video Model With 2K Output and Native Stereo Audio

Introduction

MiniMax has officially released MiniMax H3, a general-purpose multimodal video-generation model designed to understand text, images, video, and audio within one shared generation workflow.

The launch marks a shift away from treating text-to-video, image-to-video, reference-video generation, audio synchronization, and editing as isolated tasks. Instead, H3 is designed to take multiple types of context together and use natural-language instructions to describe how those references should interact.

According to MiniMax, H3 supports native stereo audio-video output, generates clips of up to 15 seconds, and provides 2K output through its current API.

The company is positioning the model for commercial content creation, including advertising, branding, e-commerce, product design, UI/UX, and gaming.

The release also puts a strong emphasis on generation cost. MiniMax says H3 was designed to deliver a lower cost per second than many mainstream video-generation systems while still supporting multimodal references and high-resolution output.

H3 Moves From Specialized Video Tasks to Unified Multimodal Generation

Earlier generations of video models were often organized around separate workflows.

A creator might choose one model or endpoint for text-to-video, another for image animation, and a different process for reference-video editing or audio synchronization.

H3 attempts to combine these use cases around a shared multimodal context.

MiniMax’s current API documentation describes three main generation modes:

The model is therefore not limited to turning a sentence into a video.

It can also use reference media as part of the generation context.

For reference generation, MiniMax’s API supports combinations of:

• Up to 9 reference images
• Up to 3 reference videos
• Up to 3 reference audio clips
• Up to 12 media files in total

Reference videos can total up to 15 seconds, while audio must be accompanied by at least one image or video reference.

This structure allows a creator to describe relationships between different inputs instead of manually handling each modality as an independent stage.

Natural Language Becomes the Bridge Between Modalities

The AIBase source describes one of H3’s key ideas as Contextual Omni Representation.

The concept is to use natural language as the interface that connects different types of media.

A user can provide several references and describe the intended relationship in ordinary language.

For example, a workflow might conceptually ask H3 to:

• Use the camera movement from one reference video
• Preserve the person shown in a reference image
• Follow the rhythm or voice from an audio reference
• Apply all of those constraints to one newly generated scene

That is substantially different from a simple text-to-video prompt.

The model has to understand not only what each media input contains, but also what role each input should play in the final generation.

MiniMax’s public API reflects this design through role-based multimodal inputs such as:

• first_frame
• last_frame
• reference_image
• reference_video
• reference_audio

The user still supplies a text prompt, but the prompt can now act as an instruction layer over several media sources.

H3 Supports Native Stereo Audio and Up to 15-Second 2K Output

MiniMax says H3 can generate video with native stereo audio rather than requiring a separate audio-generation pipeline afterward.

The current developer documentation lists:

• Model name: MiniMax-H3
• Output resolution: 2K
• Output duration: up to 15 seconds
• Common aspect ratios: supported
• Adaptive aspect ratio: available for suitable reference workflows

The API reference currently accepts integer durations from 4 to 15 seconds, while the higher-level developer guide describes the normal output range as 5 to 15 seconds.

That small documentation difference is worth noting when building against the API: developers should follow the current endpoint schema used by their account and SDK.

For text-only generation, the aspect ratio must be supplied explicitly.

Supported ratios in the current API reference include:

• 21:9
• 16:9
• 4:3
• 1:1
• 3:4
• 9:16

Reference workflows can also use adaptive sizing.

Early Testing Focuses on Commercial Content Creation

MiniMax says early testing showed strong performance in several practical areas.

These include:

• Instruction following
• Brand and text presentation
• Video-to-video motion transfer
• Multimodal generation
• Controlled editing
• Commercial creative production

The company specifically highlights use cases in:

• Advertising
• Branding
• E-commerce
• Product design
• UI/UX
• Gaming

These performance descriptions come from MiniMax and the launch coverage rather than from an independent public benchmark.

That distinction matters.

Video-generation quality is difficult to reduce to one score. A model may be strong at motion transfer but weaker at fine facial detail, typography, long-term identity consistency, or complex multi-object interaction.

The strongest way to evaluate H3 for production is therefore to test it against the actual content a team needs to create.

H3 Uses a New Multimodal Technical Stack

The AIBase report and MiniMax’s launch materials describe several technologies behind H3:

• Contextual Omni Representation
• H3-VAE
• H3-Omni Transformer
• In-Context Regeneration

The public API documentation currently focuses more on how to use H3 than on publishing a complete research paper for these components, so detailed architectural claims should be treated as MiniMax’s product description unless a technical report provides additional reproducible detail.

Contextual Omni Representation

This component is intended to represent text, image, video, and audio together in a shared context.

Its role is to help H3 understand how several reference sources relate to a natural-language instruction.

H3-VAE

H3-VAE is described as part of the model’s video representation and generation stack.

Video VAEs typically compress high-dimensional video information into a more manageable latent representation for generation and decoding.

MiniMax has not yet published enough public technical detail in the documentation reviewed for this article to independently characterize the exact design of H3-VAE.

H3-Omni Transformer

The H3-Omni Transformer is presented as the model component responsible for unified multimodal generation.

Its naming reflects the model’s broader goal: one system should reason over multiple media types instead of switching among independent specialist models.

In-Context Regeneration

MiniMax also lists In-Context Regeneration as part of the H3 stack.

The intended role is to improve generation by using information already present in the multimodal context.

As with the other architecture names, detailed training procedures and ablation results were not available in the public API documentation reviewed at publication time.

H3 Pricing Is Based on Video Duration

The AIBase article emphasizes H3’s price-performance ratio.

MiniMax’s current official pay-as-you-go pricing gives a clearer numerical reference.

Input reference materials have separate rules.

MiniMax currently lists:

• Audio references: free
• First 5 image references: free
• Additional image references: $0.04 each
• Video references: billed according to input-video duration and target output resolution

MiniMax says that, on a per-second basis, H3 at 2K costs less than one-third of mainstream competing models and that 768P costs less than half the price of mainstream 720P alternatives.

Those relative comparisons are vendor claims because the result depends heavily on which competing models, plans, resolutions, and billing methods are included.

For budgeting, developers should rely first on the published H3 API rate and then compare it with the exact alternatives they actually use.

Why Cost per Second Matters for AI Video

Video generation can become expensive quickly because cost scales with output duration.

That changes the economics of iteration.

A creator rarely generates one perfect clip.

A production workflow may involve:

1. Several prompt experiments
2. Multiple character-consistency attempts
3. Camera-motion variations
4. Different aspect ratios
5. Brand or product corrections
6. Final high-resolution generations

Even a modest reduction in cost per generated second can become significant when a team creates dozens or hundreds of alternatives.

This is especially relevant in advertising and e-commerce, where one campaign may need:

• Multiple products
• Several markets
• Different languages
• Horizontal and vertical formats
• Several creative variants

H3’s price positioning is therefore aimed as much at iteration economics as at the cost of a single final clip.

Reference Generation Is One of H3’s Most Important Capabilities

The current MiniMax API supports reference generation using a mixture of images, videos, and audio.

That allows workflows such as:

• Keeping a product or character consistent from an image
• Following the motion from a reference video
• Borrowing camera behavior from another clip
• Using a reference audio track for timing or voice
• Combining several forms of reference material in one request

MiniMax says H3 can preserve features from the reference subject or asset throughout the generated output.

This is especially relevant to commercial work.

A general text prompt can produce a visually attractive clip while still changing:

• A logo
• Product proportions
• Clothing
• Character identity
• Packaging
• Brand colors

Reference-driven workflows provide more control over those variables.

They do not guarantee perfect identity preservation, so teams should still review each generated asset before publication.

First- and Last-Frame Control Adds Another Production Workflow

H3 also supports using a first frame, a last frame, or both.

This is useful when a creator already knows how the shot should start or finish.

Typical uses include:

• Animating a product still
• Creating a transition between two images
• Matching the beginning of one shot to the end of another
• Controlling the end state of a transformation
• Building more predictable editing sequences

MiniMax’s API treats first/last-frame generation and reference generation as separate modes.

A request cannot mix first_frame or last_frame roles with reference_image, reference_video, or reference_audio roles in the same task.

That constraint is important for developers integrating the API.

MiniMax Plans to Release H3 Model Weights

One of the most notable parts of the announcement is MiniMax’s plan to make H3’s model weights available.

The company says the weights are planned for release in the coming days, subject to applicable laws and regulations.

MiniMax argues that releasing the weights can help:

• Expand the open model ecosystem
• Support customized versions
• Accelerate adaptation to different hardware
• Reduce dependence on one hosted service
• Encourage research and deployment experimentation

The company also says hardware compatibility was considered from the beginning of H3’s design, including compatibility with a broader range of AI hardware.

At the time this article was prepared, MiniMax’s official GitHub and Hugging Face organizations were publicly available, but a confirmed public H3 weights repository was not yet linked from the official H3 API documentation reviewed here.

That means developers should wait for the official MiniMax release link rather than downloading weights from an unofficial mirror.

A Correction to the “First Open-Weight Chinese Video Model” Claim

The AIBase article describes H3’s planned release as the first time a Chinese video-generation foundation model has opened its weights to the community.

Taken broadly, that statement is not historically accurate.

Alibaba released weights and inference code for its Wan2.1 video-generation models in February 2025, with later Wan2.1 variants also released publicly.

H3’s more defensible point of distinction is its general-purpose unified multimodal video architecture, including text, image, video, and audio references plus native audio-video output—not that it is the first Chinese video model with open weights.

Open Weights Could Help Hardware Adaptation

Open model weights can matter in markets where organizations want more control over infrastructure.

A hosted API is easier to start with, but open weights can make it possible to:

• Deploy on private infrastructure
• Optimize kernels for local hardware
• Quantize the model
• Build specialized inference runtimes
• Fine-tune or adapt components where licensing permits
• Integrate with domestic accelerators
• Keep sensitive data inside a controlled environment

Whether H3 can be practically self-hosted will depend on information that is not yet public in the source article, including:

• Parameter count
• VRAM requirements
• Inference precision
• Parallelism requirements
• Minimum hardware configuration
• License terms
• Training or fine-tuning support

Until the official weights and technical documentation are released, claims about consumer-GPU compatibility or self-hosting cost would be speculative.

MiniMax Acknowledges That H3 Still Has Room to Improve

MiniMax also says the model is not the final endpoint of the H series.

The company identifies areas including:

• Fine image detail
• Overall model scale
• Further capability expansion

MiniMax says it plans to continue scaling H-series models and eventually combine capabilities from the H and M model families.

The H family is currently focused on multimodal generation, especially video.

The M family includes MiniMax’s language and agent models.

A future convergence could point toward systems in which one model family handles:

• Language reasoning
• Agent execution
• Image understanding
• Video understanding
• Audio
• Video creation
• Editing

That remains a forward-looking direction rather than a currently released unified product.

What H3 Changes for Video-Generation Workflows

H3’s most important contribution is not simply higher resolution.

The larger shift is that several previously separate creative operations can now be described in one context.

A creator can move from:

Generate a video from this sentence.

toward:

Use this person, follow the motion in this video, preserve this visual identity,
take timing cues from this audio, and create a new scene described by this prompt.

That changes the role of the video model.

It becomes less like a one-purpose generator and more like a multimodal creative engine.

For commercial teams, the value will depend on how consistently H3 can preserve references, follow complex instructions, render brand information, and remain economically efficient across repeated generations.

常见问题

What is MiniMax H3?

MiniMax H3 is a general-purpose multimodal video-generation model from MiniMax. It accepts text, images, video, and audio as context and supports text-to-video, image-to-video, reference-based generation, and controlled video creation.

What resolution does MiniMax H3 support?

MiniMax’s current API documentation lists 2K as the primary H3 output resolution. Its pricing page also lists a 768P billing tier.

How long can MiniMax H3 videos be?

The official H3 documentation supports video output up to 15 seconds. The API reference currently accepts integer duration values from 4 through 15 seconds.

Does MiniMax H3 generate audio?

Yes. MiniMax describes H3 as supporting native stereo audio-video output, so audio can be generated as part of the video workflow rather than always requiring a separate post-production model.

How much does the MiniMax H3 API cost?

MiniMax currently lists H3 at $0.13 per generated second for 2K and $0.09 per second for 768P. Reference videos and additional images can add input-material charges under the published billing rules.

Can H3 use reference images, videos, and audio together?

Yes. The official API supports reference images, videos, and audio in the same reference-generation workflow, subject to file-count, duration, size, and input-combination limits.

Is MiniMax H3 open source?

MiniMax has announced plans to release H3 model weights in the coming days, subject to applicable regulations. At the time this article was prepared, the official API was live, but a confirmed public H3 weights repository was not yet linked from the official documentation reviewed here.

Is H3 the first Chinese open-weight video model?

No, not in the broad historical sense. Alibaba released Wan2.1 video-generation model weights in 2025. H3 is more notable for combining multimodal references and native audio-video generation within one general-purpose video model.

相关工具

• MiniMax H3 API: Official documentation for H3 text-to-video, image-to-video, and reference-generation workflows.
• MiniMax Open Platform: MiniMax’s developer platform for API keys, model access, billing, and developer resources.
• MiniMax GitHub: The company’s official GitHub organization for public code and model-related projects.
• MiniMax Hugging Face: MiniMax’s official Hugging Face organization for public model resources.
• Wan2.1: Alibaba’s open video-generation model family, useful as historical context for open-weight video models.

Related Links

• MiniMax Official Website: MiniMax’s official site, currently listing H3 as its new video-generation model.
• MiniMax H3 Video Generation Guide: Official model specifications, supported modes, input limits, and workflow documentation.
• MiniMax H3 API Reference: The official V2 endpoint specification for creating H3 video-generation tasks.
• MiniMax API Pricing: Current official pay-as-you-go prices for H3 and its reference-media billing rules.
• MiniMax Official GitHub: The official organization to monitor for released code and model resources.
• Alibaba: Wan2.1 Open-Source Video Models: Official historical confirmation that Chinese open-weight video models were available before H3.
• Wan2.1 GitHub Repository: The public inference code and weights for Alibaba’s earlier open video-generation family.

Summary

MiniMax H3 brings text, images, video, and audio into one general-purpose video-generation workflow. It supports up to 15-second 2K output, native stereo audio, first/last-frame control, and reference generation using multiple types of media.

Its commercial pitch centers on two things: control and cost. MiniMax says H3 performs well on instruction following, brand presentation, and motion transfer, while its official API currently lists generation at $0.13 per second for 2K and $0.09 per second for 768P.

MiniMax also plans to release the model weights, although a confirmed H3 weights repository was not yet linked in the official documentation reviewed at publication time.

H3’s most important shift is not simply 2K video—it is the move toward one multimodal generation system that can understand how text, images, video, and audio should work together.