Description
Running video generation yourself usually hits two walls: hosted services bill by the second and want your footage on someone else's server, and open models tend to give you pictures with no sound, or sound you have to bolt on afterwards. MiniMax H3 closes both gaps — the open-weights H3-Base produces 768p video and synchronized stereo audio in a single pass, so a character's lip movement and voice line up without a separate dubbing step.
The full weights sit on Hugging Face and ModelScope: a 33B-parameter single-stream Transformer. If you want a visual workflow, ComfyUI ships native text-to-video, image-to-video and reference-to-video templates; to serve it to a team, the official route is a self-hosted SGLang service, with a community recipe for vLLM-Omni. Note that this is open weights, not the whole product: the H3-Context-IR instruction reader and the 2K regeneration stage are not released, so a local run gives you the 768p base, not the full pipeline of MiniMax's own platform.
Text to audio-video: one prompt yields 4–15 seconds at 24 fps plus 32 kHz stereo audio, with dialogue in 11 languages including English, Chinese and Japanese.
First and last frame: give the first frame, the last frame or both, and the model fills in motion and sound between them — handy for turning storyboard stills into shots.
Reference generation: feed up to 9 images, 3 video clips and 3 audio clips to pin down faces, scene style and voice; this is a separate Ref2VA checkpoint.
Several ways to deploy: a local ComfyUI workflow (0.30.0 or later), an async /v1/videos API on SGLang, or vLLM-Omni serving both checkpoints from one service.
Heavy hardware: the reference setup is 8 × B200; the community has run it on a single 24 GB RTX 4090 with offloading and INT8, but plan for 100 GB+ of disk and, in some recipes, 200 GB+ of RAM.
Check the license: it ships under the MiniMax H3 Community License; the US, EU, UK and South Korea are excluded territories that need a separate license, so read it before commercial use.
The full weights sit on Hugging Face and ModelScope: a 33B-parameter single-stream Transformer. If you want a visual workflow, ComfyUI ships native text-to-video, image-to-video and reference-to-video templates; to serve it to a team, the official route is a self-hosted SGLang service, with a community recipe for vLLM-Omni. Note that this is open weights, not the whole product: the H3-Context-IR instruction reader and the 2K regeneration stage are not released, so a local run gives you the 768p base, not the full pipeline of MiniMax's own platform.
Features
Text to audio-video: one prompt yields 4–15 seconds at 24 fps plus 32 kHz stereo audio, with dialogue in 11 languages including English, Chinese and Japanese.
First and last frame: give the first frame, the last frame or both, and the model fills in motion and sound between them — handy for turning storyboard stills into shots.
Reference generation: feed up to 9 images, 3 video clips and 3 audio clips to pin down faces, scene style and voice; this is a separate Ref2VA checkpoint.
Several ways to deploy: a local ComfyUI workflow (0.30.0 or later), an async /v1/videos API on SGLang, or vLLM-Omni serving both checkpoints from one service.
Heavy hardware: the reference setup is 8 × B200; the community has run it on a single 24 GB RTX 4090 with offloading and INT8, but plan for 100 GB+ of disk and, in some recipes, 200 GB+ of RAM.
Check the license: it ships under the MiniMax H3 Community License; the US, EU, UK and South Korea are excluded territories that need a separate license, so read it before commercial use.
Tags:video-generation