MiniMax drops H3: A 2K video and audio open model

Anyone who has pieced together an AI video knows the drill: generate a still in one app, cross your fingers animating it in another, and head to a third tool for sound effects and background music. If you need a camera cut to land on the beat, you almost have to start from scratch.

MiniMax launched H3, an omni-modal foundation model built to end that fragmented pipeline. The company behind Hailuo AI is not just bundling text, images, video, and stereo audio into a unified architecture. It is also promising to open-source the model weights in the coming days, taking a rare open-weights swing at a space dominated by closed corporate labs.

Four modalities under one prompt

H3 natively handles text, images, video, and audio in a single context window, producing up to 15-second clips at 2K resolution complete with native stereo sound.

Traditional video models divide tasks into specialized silos: text-to-video, image-to-video, motion transfer, first-and-last frame, and audio generation. MiniMax discarded that patchwork design. H3 lets natural language act as the universal director across modalities.

You can give it a prompt like: "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, and sync the vocals to Audio 3." The model parses the multimodal inputs and resolves the relationships itself.

Early hands-on testing by creators shows this flexibility in action. In one demo, an independent reviewer fed H3 two storyboard images and a brand logo to generate a luxury handbag ad. In another, H3 took an audio track and a character reference image, delivering a music video with hard cuts timed within three seconds of the underlying beat.

Ditching the super-res crutch

To make full-modality generation practical, MiniMax overhauled its underlying pipeline.

Handling multimodal inputs typically blows up sequence length. H3 runs raw reference material through a specialized understanding pipeline that condenses roughly 100K inference tokens down to an average of 4K tokens, creating what the team calls Contextual Omni Representation.

Architecturally, MiniMax set aside its Hailuo 02 framework in favor of the H3-Omni Transformer. By decoupling compute workloads for understanding and generation, the team lifted training throughput by nearly 30%.

Its 2K output also skips the conventional super-resolution module. Dedicated upscalers usually guess missing pixels, often warping typography and fine logos. H3 instead uses In-Context Regeneration: the base model regenerates its own lower-resolution output using the original multimodal prompt, preserving crisp text and micro-details.

MiniMax claims H3's per-second price at 2K is less than one-third that of mainstream commercial video models. In hands-on testing, a 10-second generation cost 120 credits on its web platform—about $1.20, which the reviewer noted was roughly three times cheaper than competing commercial services like SeaDance 2.5.

Why it matters

For creators and developers, H3 could be a massive release if the open weights deliver on their promise. High-end video generation with prompt-accurate native audio has remained locked inside closed commercial APIs. An open-weights release means studios and independent developers can fine-tune custom pipelines for e-commerce, advertising, and game assets without recurring API bills or platform restrictions.

For the broader AI ecosystem, MiniMax is shifting the battleground from isolated video generation to unified multimodal synthesis. Generating synchronized stereo audio and video inside a single model bypasses messy multi-tool workflows and raises the bar for competitors still relying on chained specialized models.

We think betting on natural language as the universal bridge across tasks is the right technical direction. It treats multimodal generation as an integrated intelligence problem rather than a collection of separate features. That said, MiniMax acknowledged that visual detail still has room to improve in complex scenes. The real test will come when the weights drop: how easily H3 runs on community hardware will determine whether it truly shakes up the video generation market.