What local deployment includes
The open release is H3-Base, the video-and-audio generator at the center of the complete H3 system. It accepts a prompt and optional visual or audio conditions, then produces 24 fps video with a 32 kHz stereo soundtrack. The native output recipe uses a 768-pixel short edge, such as 1344 x 768 for 16:9.
Two surrounding modules are different. H3-Context-IR interprets complex multimodal instructions before generation, and H3-Regenerate-2K regenerates a selected 768p result at higher resolution using the original context. MiniMax says both modules are hosted and are not included in the initial open release.
| Module | Local weights | Job |
|---|---|---|
| H3-Context-IR | No | Hosted instruction processing for complex multimodal requests |
| H3-Base | Yes | Generates 768p video and stereo audio |
| H3-Regenerate-2K | No | Regenerates a selected base result at 2K |
Choose the checkpoint before the framework
FL2VA and Ref2VA are separate task-specific checkpoint families. Downloading both wastes storage when a machine serves only one workflow.
- FL2VA: text-to-video, first-frame-to-video, last-frame-to-video, and first-and-last-frame generation.
- Ref2VA: reference-driven generation with up to 9 images, 3 videos, and 3 audio clips within the documented combined limits.
- Video-to-video is a Ref2VA use case, not a separate public checkpoint or task value.
Pick a deployment path
ComfyUI 0.30.0 and later includes native H3 nodes and templates. SGLang publishes measured recipes for NVIDIA datacenter cards, dual RTX 5090, and AMD Instinct hardware. Diffusers can fetch only the components required by its modular pipeline. Avoid mixing commands, model layouts, or quantization files from different frameworks unless the relevant guide explicitly supports that combination.
| Path | Best for | Trade-off |
|---|---|---|
| ComfyUI | Workstations and visual iteration | Fastest setup; hardware-specific repacks and offloading vary |
| Diffusers | Python experiments and custom pipelines | Direct control; you own memory placement and serving |
| SGLang | Measured multi-GPU serving | Strong official recipes; production hardware is expensive |
| vLLM | API-oriented serving teams | Familiar service workflow; follow the current H3 recipe exactly |
Validate with a small reproducible run
Do not start with a 15-second reference-heavy job. First prove that model loading, conditioning, video decode, audio decode, and file export all work on the target machine.
- Use the 768-pixel-short-edge recipe only after a smaller preview completes.
- Fix the prompt, seed, dimensions, duration, step count, and model revision.
- Check the MP4 for video, stereo audio, duration, frame size, and a conditioning input that visibly affects the result.
- Record peak VRAM, peak system RAM, model load time, denoise time, decode time, and end-to-end time.
The practical recommendation
Use ComfyUI when the goal is local creative iteration on a single workstation. Use SGLang or vLLM when several users need a stable endpoint and the deployment has multiple large GPUs. Use the local base model to test prompts and motion at 768p, then send only selected clips through the hosted 2K stage when the project permits an API boundary.