MiniMax H3 dropped today with open weights and is natively supported in ComfyUI as of this morning, delivering day‑zero access to a next‑generation open‑weights video model. The system accepts text, images, video, or audio and generates video with real stereo sound, up to 2K resolution and 15 seconds per clip. As MiniMax’s third‑generation model, H3 follows Hailuo 01 and Hailuo 02 and is the first with open weights.

Users can try text‑to‑video prompts, image‑to‑video, first‑and‑last‑frame control, and reference‑to‑video, where a reference image, video, or audio carries a subject, motion, or voice through the clip. Audio is generated natively in stereo during the same pass, eliminating post‑processing. A key breakthrough is motion transfer for graph work, allowing reference videos to supply camera moves or performance while style and subject come from elsewhere.

To run locally, MiniMax pruned ~40 % of modulation weights, replaced them with lookup tables, and applied int8 quantization with custom kernels, shrinking the memory footprint by 66 % to 42.5 GB. Dynamic VRAM offloading lets the model run on consumer GPUs like RTX 3060. Update ComfyUI to version 0.30.0, download the I2V, R2V, or T2V workflows, and start creating.