MiniMax H3
MiniMax H3 from MiniMaxAI: text-to-video and image-to-video; Planned.
MiniMax H3 is a video generation model from MiniMaxAI. The catalog records it for text-to-video, image-to-video, and reference-to-video. It is also cataloged under the aliases minimax_h3, minimax-h3-t2va, and minimax-h3-fl2va.
No official teaser or method figure is checked into this docs tree.
Code: sgl-project/sglang.
Build your run
NativeMiniMaxH3Pipelineworldfoundry.pipelines.minimax.pipeline_minimax_h3defaultworldfoundry.pipelineworldfoundry-eval zoo model-download --model-id minimax-h3 --check-local --jsonWhat it is for
minimax-h3 is the recorded prompt-to-video route — not image editing and not a policy. Start here when the job is a clip from text.
Catalog tasks include text-to-video, image-to-video, and reference-to-video. Catalog aliases include minimax_h3, minimax-h3-t2va, minimax-h3-fl2va, and minimax-h3-ref2va. Documented modalities: text and image in, video out.
Runtime contract
Optional prompt, steps, seed, task, and short_edge; artifacts minimax-h3.mp4.
CLI task profiles: default. WorldFoundry binds NativeMiniMaxH3Pipeline.
Family status: Planned.
Joint audio+video generation; 32kHz stereo audio track muxed into the mp4.
Ported faithfully from SGLang multimodal_gen (Apache-2.0), single-GPU PyTorch.
Running in WorldFoundry
The catalog launch is worldfoundry-eval run minimax-h3 --pipeline.task-profile default. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.
It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).
Recorded weight repositories include MiniMaxAI/MiniMax-H3. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh minimax-h3. Add --download to fetch missing public checkpoints.
Recorded licenses include apache-2.0; catalog VRAM 80 GB; status: Planned.
