MiniMax H3 is a 33-billion-parameter video generation model that produces up to 2K resolution video with synchronized stereo audio, and unlike competitors Sora, Veo, and Kling, it's released with open weights so teams can run it themselves. The model accepts text, images, video clips, and audio as input and can generate videos between 4 to 15 seconds long at 24 frames per second in multiple aspect ratios, from ultra-wide 21:9 to vertical 9:16 formats.

Most AI video models that dominated headlines this year operate as closed black boxes. You send a prompt to an API, receive a video clip back, and have no visibility into how the system works beyond what the company's marketing materials reveal. MiniMax H3 breaks that pattern by publishing its architecture, releasing model weights that anyone with sufficient computing power can download and run, and providing inference recipes for popular frameworks like SGLang, vLLM, diffusers, and ComfyUI.

What Makes MiniMax H3 Different From Closed Video Models?

The distinction between open and closed models matters more than raw capability benchmarks. With H3, teams gain transparency into exactly how the system processes inputs and generates outputs. They can audit the model for bias, customize it for specific use cases, and avoid dependency on a third-party API that could change pricing, availability, or terms at any time. For enterprises building video generation into their products, this control is significant.

H3 supports an unusually flexible input system called "omni-reference" mode, which accepts up to nine images, three video clips, and three audio clips simultaneously. Rather than working from a single text prompt and maybe one reference image like most generators, H3 can synthesize a new video that draws style, subjects, and sound from all of them at once. This mirrors how creative briefs actually get handed to human video editors, making the tool feel less like a text-to-image toy and more like a professional editing assistant.

How Does the Three-Stage Architecture Work?

H3 isn't a single monolithic model. It's three components that run in sequence, and understanding the split reveals both its strengths and what remains proprietary.

• H3-Context-IR (Instruction Representation): This preprocessing layer parses messy inputs, figures out how text, images, video, and audio relate to each other, handles temporal understanding, and fills in underspecified details. It's not open-sourced because it depends on multiple internal models, but MiniMax publishes prompting guidance so teams can build an equivalent layer themselves.
• H3-Base (The Open Core): This is the 33-billion-parameter transformer that teams can download and run. It uses a single-stream architecture where text, image, and audio all flow through the same transformer blocks rather than separate expert pathways. Two separate VAEs (variational autoencoders) compress video and audio into latent tokens that the transformer jointly predicts, producing synchronized sound rather than audio bolted on afterward.
• H3-Regenerate-2K (Upscaling): Rather than using a conventional super-resolution model, H3 feeds its low-resolution output back into itself alongside the original context and regenerates at 2K resolution. This lets the model recover detail it already understood rather than hallucinating fine details from blurry pixels.

The practical result is that the generation core is self-hostable, while the orchestration layers currently require MiniMax's hosted API. This is a meaningfully different position than a fully closed model, and a meaningfully different position than a fully open one.

What Are the Hardware Requirements for Running H3?

MiniMax's reference deployment targets four GPUs with tensor and sequence parallelism, a meaningful but not exotic hardware bar. This is closer to what a mid-sized team already running open-weight large language model inference would have on hand than to a hyperscaler-only requirement. The model ships as two task-specific checkpoints: H3 Base FL2VA for text-to-video and first/last-frame-to-video tasks, and H3 Base Ref2VA for omni-reference generation.

Both checkpoints are distributed as self-contained repositories that can be pulled with a scoped Hugging Face download, loaded directly through diffusers' ModularPipeline, or served with SGLang or vLLM. For teams that want the full 2K pipeline end-to-end without standing up H3-Context-IR themselves, MiniMax exposes a combined API workflow that uses locally hosted H3-Base for generation while handling preprocessing and upscaling through hosted services.

How to Deploy MiniMax H3 for Your Team

• Evaluate Your Hardware: Assess whether your infrastructure has access to four GPUs or can be scaled to that level. If you're already running open-weight language models, you likely have the foundation in place.
• Choose Your Integration Path: Decide whether you want full self-hosting with the open-source H3-Base model, or a hybrid approach using MiniMax's hosted Context-IR and Regenerate-2K services alongside your local generation core.
• Select the Right Checkpoint: Pick H3 Base FL2VA if you're primarily doing text-to-video and frame-conditional generation, or H3 Base Ref2VA if you need the omni-reference capability to blend multiple images, video clips, and audio sources.
• Integrate With Your Framework: Load the model through diffusers, SGLang, vLLM, or ComfyUI depending on your existing pipeline and team expertise.

The release of H3 with open weights signals a shift in how the video generation industry is stratifying. While frontier labs like OpenAI, Google, and ByteDance continue to keep their models closed, MiniMax is betting that transparency and self-hosting capability will appeal to enterprises that need control, auditability, and independence from API providers. For teams evaluating video generation tools, the choice between closed APIs and open models now has a credible third option.