MiniMax H3 Max R2V

• 👁️ Reference-to-video: image, video and audio references steer generation
• 🏢 The "Max" tier of MiniMax's H3 video model line
• 📏 Documented output resolutions ranging from 768P up to 2K
• ⚡ Clip durations of roughly 4 to 15 seconds
• 🔧 Aspect ratio defaults to adaptive unless explicitly set
• 🎯 Same endpoint also serves text-to-video and image-to-video
• 🌐 Accessed through MiniMax's video generation V2 API
• 🆕 Released August 2026 alongside H3 Max text- and image-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 Max R2V is the reference-to-video configuration of MiniMax's H3 Max video model, released in August 2026 alongside its text-to-video and image-to-video counterparts, MiniMax H3 Max and MiniMax H3 Max. In MiniMax's own API documentation, H3 and H3 Max appear as selectable model entries on the same video generation V2 endpoint, which accepts multimodal input — text, image, video and audio — and covers text-to-video, image-to-video with frame-level control, and reference-to-video, at resolutions from 768P up to 2K and durations of roughly 4 to 15 seconds.

Reference-to-video is the mode in which generation is conditioned on supplied material rather than a prompt alone. Requests carry reference images, reference videos and reference audio, and the aspect ratio is adaptive by default unless a value is specified. This makes the mode suited to work where a subject, look or sound has to be carried across shots rather than re-described in text each time.

Relative to the earlier H3-generation reference models — MiniMax H3 R2V from July 2026 and MiniMax H3 R2V Enhanced from earlier in August — the Max entry is the higher-capability tier exposed on the same endpoint and shares its request format and reference-conditioning options. MiniMax has not published detailed public metrics separating the Max tier from those predecessors, so this listing stays limited to documented capabilities rather than performance claims.