MiniMax H3 in Karya: 2K Video with Native Audio and Reference Locking

MiniMax unveiled H3 (also known as Hailuo 03 or Hailuo 3.0) at WAIC 2026 and released it on July 31, 2026. It is MiniMax's next-generation multimodal video model - a step up in both resolution and how the model understands its inputs, and it lands in Karya's Video Studio and Flow the same week.

What is new in H3

• Clips from 4 to 15 seconds at up to 2K (2560x1440) resolution, 24fps.
• Native stereo audio generated together with the video in one pass, instead of a separate audio step bolted on afterward.
• A single unified model that understands text, images, video and audio as inputs, not just a text prompt.
• Strong instruction following, so detailed prompts about action, camera and mood carry through into the result.

Reference mode: lock identity, style, motion or voice

H3's Reference mode accepts up to 9 reference images, 3 reference video clips and 3 reference audio clips in a single generation. That is enough to lock a character's face and outfit, match a visual style, carry a specific motion from one clip onto a new scene, or keep a voice consistent - useful for a recurring character across an ad series, or a product that has to look identical shot to shot.

An open-weight promise

MiniMax has said it plans to release H3's weights openly, which would make it the first frontier video model with an open-weight release. That is a stated intention rather than a shipped download today, but it is a notable direction for a model at this quality level.

Using MiniMax H3 in Karya

Open the Video Studio and pick "MiniMax H3" for text-to-video, or "MiniMax H3 (Reference)" to feed in reference images, video and audio. In Karya Flow, three MiniMax H3 nodes cover text-to-video, first/last-frame image-to-video, and reference-to-video - chain any of them behind a prompt or LLM text node, or fan a single reference set out across a List node to batch several variations from one setup.

Pricing

MiniMax H3 is billed at roughly 32 credits per second of video - a default 6-second clip works out to about 192 credits. Karya shows the exact credit cost before you generate, you only pay for generations that complete, and unused credits never expire. New accounts start with 200 free credits.