MiniMax H3 R2V Enhanced

• 🆕 Reference-to-video variant of MiniMax's H3 (Hailuo 3.0) generation family.
• 🆕 Released August 5, 2026; newest model in this reference-to-video line.
• 👁️ Takes reference visual material alongside a text prompt to guide output.
• 🧠 "Enhanced" path is associated with MiniMax's multimodal context interpretation task.
• 💬 MiniMax API docs describe an H3-Context-IR task returning an enriched video prompt.
• 📏 Community write-ups describe H3 as generating video with natively synchronized audio.
• 🌐 H3 is published as open weights; quantized community builds exist on Hugging Face.
• 🔧 Siblings cover text-to-video, image-to-video and non-enhanced reference-to-video.

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax H3, also branded Hailuo 3.0, is MiniMax's multimodal video generation model, described in community documentation on Hugging Face as an open-weight system that returns video together with natively synchronized audio rather than a silent track. This catalog entry covers the reference-to-video (R2V) task: instead of describing a scene from scratch, you supply reference material and a prompt so that identity, style and motion characteristics carry through into the generated clip.

Relative to its same-family predecessor MiniMax H3 R2V, the non-enhanced reference-to-video sibling, this August 2026 model keeps the same reference-driven task but routes requests through MiniMax's enhanced-prompt path. MiniMax's API documentation describes an H3-Context-IR task that interprets multimodal context — text together with visual and audio material — and returns only an enriched video prompt for use in generation. Based on that documentation, the enhanced path appears to add a prompt-interpretation step ahead of generation, though MiniMax's docs do not spell out the architectural relationship.

Its text-driven counterpart is MiniMax H3 Enhanced, while MiniMax H3 text-to-video and MiniMax H3 image-to-video cover the plain prompt-only and single-image entry points. Outside the video line, MiniMax also ships the MiniMax M3 Preview text model and audio systems such as MiniMax Speech-02 HD and MiniMax Music 2.6.

Because H3 weights are openly published, third-party quantized builds of the base model are available on Hugging Face for local inference.