MiniMax Releases MiniMax H3 Omni-Modal Video Model

MiniMax H3 Folds Video, Audio, and Images Into One Model

MiniMax H3 is here, and it is not another incremental video upgrade. The company released its general-purpose multimodal generation model on July 31, 2026, and it is live in the platform API under the model ID MiniMax-H3, plus the consumer Hailuo AI app. This is a single system that reads text, images, video, and audio as one unified context and returns video with native stereo sound. It outputs 2K resolution clips that run between 4 and 15 seconds, with integer durations only.

The old approach to video generation was a mess of specialized tools. You had one model for text-to-video, another for image-to-video, a third for first-and-last-frame interpolation, plus separate systems for subject reference, motion reference, and video editing. MiniMax H3 collapses all of that into one pretraining paradigm. Reference and editing relationships are expressed in natural language. The example prompt from MiniMax makes the intent clear: reference the camera movement from Video 1, have the character in Image 2 sing, match the vocals to Audio 3.

That is a fundamentally different way to think about generation. Instead of telling a model which task you want, you describe the relationship between assets. Language becomes the bridge that turns a fixed task set into an open, descriptive one.

What the API Actually Accepts

Deployment is a mixed bag today. The API is live and working, but open weights are promised "in the coming days," not shipped. If you want to run this on your own hardware, you cannot. The only path is the platform API.

The video generation guide documents three entry modes: text-to-video, first/last-frame image-to-video, and reference generation. All of this runs behind one endpoint with an asynchronous three-step flow: create a task, poll the task_id, then download content.url.

There are input limits worth designing around before you start building:

• Reference images: up to 9
• Reference videos: up to 3 clips, 2-15 seconds each, 15 seconds total max
• Reference audio: up to 3 clips, but audio cannot be sent without an accompanying image or video
• Mixed input caps at 12 files total
• Prompt length is capped at 7,000 characters
• Request body max is 64 MB, with URL input recommended for large assets

Per-asset file sizes are strict: video is limited to 50 MB, images to 30 MB, and audio to 15 MB. Supported formats include H.264/H.265 for video, JPG/PNG/WEBP/HEIC/HEIF for images, and WAV/MP3 for audio. These are not arbitrary restrictions. They force you to think about how you feed the model, which is a good thing when you are dealing with mixed multimodal context.

Four Technical Pieces Doing the Heavy Lifting

MiniMax rebuilt the stack from the ground up, and there are four distinct pieces that matter. The first is something they call Contextual Omni Representation. This is a rebuilt captioning system that describes the relationship between context and target video, not just the target itself. Most source material requires roughly 100K tokens of inference, which gets distilled to about 4K tokens on average. That is a massive compression of understanding.

The second piece is the H3-VAE, a full tokenizer overhaul. Its high compression ratio delivers a stated 4x gain in effective sequence length. That cuts training and inference cost, and it is the enabling technology for native 2K output. Without this, generating at that resolution would be economically unviable.

The third component is the H3-Omni Transformer. This is where MiniMax explicitly set aside the Hailuo-02 architecture. Multimodal context tripled sequence-length variance, so the training architecture separates understanding and generation workloads, tuning hardware utilization for each. The reported result is end-to-end training throughput up nearly 30%.

In-Context Regeneration Replaces Super-Resolution

The fourth piece is In-Context Regeneration, and this might be the cleverest part. Instead of bolting on a super-resolution module, the base model regenerates its own low-resolution output in-context, re-reading the original multimodal context. That is what recovers small text and fine detail that conventional upscalers guess at. For brand and product rendering, this is directly relevant because upscalers tend to mangle logos and small type.

At 2K, MiniMax claims H3's per-second price is less than a third of mainstream models; at 768p, less than half the price of mainstream 720p.

No change needed. But MiniMax's own pay-as-you-go page still listed only Hailuo 2.3 tiers at the time of writing, so treat that figure as reported, not primary. The company amplified both the launch and the pricing framing on X, signaling that cost is a central part of the pitch.

Who Is This Actually For?

MiniMax positions H3 for advertising, branding, e-commerce, product design, UI/UX, and gaming, along with film pre-visualization and retail catalog media. The application list is broad: ad variant generation, product and listing videos, animated posters, film title sequences, website hero loops, character-consistent game cinematics, and video-to-video motion transfer.

That last one matters. Video-to-video motion transfer with a single model is exactly the kind of workflow that used to require stitching together multiple expert systems. Now it is one prompt.

The unified context approach is not just about convenience. It changes what is possible. When a model can read a video for camera movement, an image for a character, and an audio clip for vocals, all in one pass, the creative workflow shifts from assembling pipelines to writing descriptions. The bottleneck becomes prompt design, not tool integration.

Where MiniMax H3 Stands in the Market

Placement data from SCMP, citing Artificial Analysis, says H3 leads in video editing while trailing Google's Gemini Omni Flash in text-to-video. In image-to-video, it sits behind both Seedance 2.0 and Gemini Omni Flash. So this is not a universal leader across every task, but it is competitive in the editing space.

For anyone building products on top of video generation APIs, the practical question is whether the unified approach justifies switching from a stack of specialized models. The answer depends on your workload. If you are doing heavy reference-based editing with multiple input types, H3's design is clearly aimed at you. If you just need straightforward text-to-video, the leaderboard suggests you have better options.

The open weights promise is the wildcard. If they ship as promised, this becomes a self-hostable option for teams that cannot send proprietary assets to a third-party API. That would change the calculus for a lot of studios and brands.

For now, the API is the only game in town, and it is a capable one. The model handles mixed inputs up to 12 files, returns native stereo audio, and does 2K resolution natively. That is a strong package, even before the open weights land.