MiniMax Releases H3 Model, Building a General-Purpose Creation System | Oasis Capital Vitality

Oasis Capital portfolio company MiniMax today officially released its next-generation multimodal generative model, MiniMax H3.

Between text, image, video, and sound, H3 no longer breaks text-to-video, motion reference, audio-video generation, and content editing into isolated tasks. Instead, it processes more complete creative intentions directly through unified multimodal contextual understanding.

A watershed moment in model competition is arriving: the race is shifting from single-modality, single-capability benchmarking toward general models that can understand complete intentions and handle complex tasks. This has been one of Oasis Capital's fundamental convictions in our continued bullishness on MiniMax.

A generative model that truly achieves scale effects needs language as its central hub to connect different modalities and tasks — completing understanding, reference, editing, and generation within a single system, and entering real production workflows with sufficient controllability and sufficiently low cost.

H3 is MiniMax's systematic push in this direction. The model supports native dual-channel audio-video generation with output up to 15 seconds at 2K resolution, can execute complex multimodal reference, content editing, and motion transfer tasks, and evolves from a point tool that "generates a piece of content" toward a general model that understands requirements, invokes context, and participates in complete creative workflows.

MiniMax H3 is a general-purpose omnimodal generative model with unified understanding of multimodal context composed of text, image, video, and sound, capable of outputting native dual-channel audio-video at up to 15 seconds and 2K resolution.

Based on early private beta feedback, MiniMax H3 demonstrates commercial-grade multi-scenario content generation capabilities, with strong performance in instruction following, text and brand information rendering, and V2V Motion Transfer (video-to-video motion migration). It enables precise, controllable multimodal content editing and generation, broadly applicable to commercial scenarios including advertising, branding, e-commerce, product design, UI/UX, and gaming.

At the same time, thanks to multiple technologies including Contextual Omni Representation, H3-VAE, H3-Omni Transformer, and In-context Regeneration, we are positioned to offer the best cost-performance ratio in the industry. We provide 2K resolution by default; at 2K resolution the model costs less than one-third per second compared to mainstream models, and less than one-half at 768P resolution.

For a long time, closed-source models have dominated the video generation field, with iteration speed and ecosystem openness lagging behind large language models and other domains. To advance the open-source community, accelerate domestic chip adaptation (H3 was designed from the outset with compatibility for multiple existing domestic chips in mind), and enable users with specific needs to customize their own versions, we plan to release the model weights in the coming days, subject to compliance with relevant laws and regulations.

H3 Model Features

Multimodal Contextual Understanding

Real creation requires integrating complex information across modalities, bringing in multimodal sources such as images, sound, and video simultaneously. Take the following shot prompt as an example: "Reference the Hitchcock dolly zoom from Video 1, make the character in Image 2 sing, with vocals referencing Audio 3." By describing the context and its relationship to the target video in natural language, H3 handles the complex omnimodal understanding work itself.

Input reference video, image:

H3 generated video: 2K performance: Native dual-channel:

H3 Application Cases

Film Title Sequences

Game UI

Dynamic Posters

Advertising & E-commerce

H3 Design Philosophy

Breaking Task Boundaries: From Specialized to General

We previously developed two generations of models: Hailuo 01 built the system from 0 to 1, while Hailuo 02 focused on improvements to core components including architectural efficiency, data quality, and scale.

During H3's design process, we recognized the limitations of past generative models in terms of task scope: image generation was often divided into separate expert models for T2I, editing, subject reference, motion reference, style reference, etc.; sound generation treated vocals, sound effects, and music as independent domains; video generation was further fragmented into text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, voice reference, video editing, and other niche functions — with clear boundaries between image, video, and audio as well.

This isolation of tasks, capabilities, and modalities restricts freedom of use and also limits model generalization from the training paradigm perspective. These two points correspond to enormous room for transformation in both application paradigms and training paradigms. Therefore, the first principle of H3's construction is the unification and generalization of tasks.

Based on this, here is a brief summary of H3's pre-training paradigm —

Data and Tasks

• Text-to-image
• Text-to-video
  • With joint audio generation; all audio output is in native dual-channel mode
  • Native multi-shot modeling
• Text-to-audio
  • No distinction between vocals, sound effects, and music; unified joint modeling
• Generalized reference and editing
  • Image-to-image reference and editing
  • Image-to-video reference and editing
  • Audio-to-audio reference and editing
  • Audio-video-to-audio-video reference and editing

In our design, generalized reference and editing means:

a. Composed entirely of real, natural data, thus possessing good data scalability;

b. Reference and editing relationships expressed through natural language, not limited to finite tasks; language (or broadly, intellectual structure) is the bridge to generalization

Architecture Choices

• The simplest, most general, and most efficient architecture possible, with everything in service of task generalization

Training Strategy

• Fuse all types of data and tasks as early as possible; appropriate mixing ratios are key

These choices lead to the same goal: equipping H3 with broad multimodal contextual understanding and generation capabilities during the pre-training stage.

Having discussed changes to the training paradigm, how is the application form of video models evolving?

We see that creators can directly describe tasks in natural language, rather than merely inputting a simple scene prompt. As multimodal understanding, instruction following, and complex task execution capabilities continue to improve, video models will be able to understand more complete creative intentions, handle more complex content requirements, and gradually evolve from "generating a video" toward truly participating in the entire content production process.

H3 Technical Choices

Here is a brief sharing of some of H3's core technologies:

Contextual Omni Representation

In training H3, the most important thing we did was enhance captioning capability.

• The introduction of multimodal context further generalizes the definition of captioning. We need not only to describe the target video, but also to describe the relationship between context and target video, and even relationships among elements within the context.
• We need to jointly describe video and audio, and the audio-video correlation under multi-shot conditions is even more complex.
• This is essentially a Contextual Omni Representation. Language serves as a generalizable connective and explanatory mechanism, unifying "tasks" in the form of open-ended descriptions — this is the root of H3's broad instruction understanding capability.
• To achieve these goals, we developed specialized models and full-modality understanding pipelines; most materials require 100K tokens of inference, ultimately yielding an average of approximately 4K tokens.

H3-VAE: Major Architectural Efficiency Gains

• The H series has maintained continuous exploration in tokenizer technology. For the H3 generation, we thoroughly revolutionized the previous tokenizer technology, achieving comprehensive improvements in reconstruction and learnability. This enables H3 to achieve competitive results in efficiency, while its high compression ratio delivers 4x sequence length benefits, dramatically reducing training and inference costs — a key technology supporting our provision of native 2K resolution.

H3-Omni Transformer: Architecture in Service of Task Generalization

• Based on H3's design philosophy that "architecture should serve tasks," generality and efficiency are the only two goals. Notably, we abandoned the Hailuo-02 architecture that once brought us significant architectural advantages, because it introduces additional complexity in a model definition centered on task generalization. We believe "task generalization" is an irreversible trend, and architectural tricks should yield to model definition.

• In H3, due to the introduction of multimodal context, sequence length variance increased by 3x, and the computational workload of understanding versus generation portions became significantly heterogeneous. We adopted a heterogeneous training architecture for understanding and generation, finely tuning hardware utilization under different workloads, and jointly considering per-sample heterogeneous computation × inter-sample load balancing, improving end-to-end training throughput by nearly 30%.

In-context Regeneration

• For H3's 2K resolution output, we did not choose a conventional dedicated super-resolution module. Instead, we use the H3 base model to perform in-context regeneration of its own lower-resolution results.
• This brings two advantages: ① The regeneration process can maximally reuse the generation capabilities already present in the H3 base model; ② The in-context form can again leverage the original multimodal contextual information for high-resolution output, thus surpassing information that traditional super-resolution approaches cannot restore through "guessing," such as small text and fine details.
• In-context Regeneration is also an expression of task generalization.

We will release a more detailed H3 Tech Report subsequently; we welcome more friends to exchange ideas with us.

Vision & What's Next

Language, image, video, and audio are widely existing natural modalities. They are tightly intertwined and constitute the fundamental medium through which humans interact with the world. The multimodal context formed by them can efficiently express broad information, and the expression and transmission of information is itself a form of productivity.

For multimodal understanding and generation, we can view language as a generalizable, scalable computational system; therefore, we believe multimodal capabilities should be tightly associated with language.

H3 still has some limitations. Here are directions we will improve in subsequent versions:

• The most important foundation of generation capability is multimodal contextual understanding capability; there is enormous room for improvement here. We will subsequently push for integration between the next version of the H series and M series model capabilities.
• Model scale limitations mean H3 still has room for improvement in the completion quality of some capabilities. Scaling is a clear direction, and we believe task generalization will give model scaling ample room to demonstrate its value.
• H3's visual fidelity in certain scenarios still needs improvement; we will continue pursuing higher resolution and more refined visual quality.