Video generation
MiniMax-H3
A multimodal model that generates video with native audio from text, first and last frames, or reference media. Four-node validation covers generation from text and from reference images and video.
01 / Model parameters
Model architecture & capabilities
MiniMax-H3 generates video with native audio from text, images, or multimodal references. FL2VA accepts text alone, a single image, or first and last frames. Ref2VA supports combinations of reference images, video, and audio. Each variant and workflow has its own resource requirements and generation times.
- Text-to-video with audio
- Short videos with first/last-frame control
- Creative work with reference images and video
- Publisher / model
- MiniMax · MiniMaxAI/MiniMax-H3
- Parameter count / architecture
- 33B Omni-Transformer dense backbone; approx. 13B AdaLN branches can be precomputed and pruned, plus a text encoder and audio/video VAEs
- Input / output
- Text, image, and multimodal reference input; video and 32 kHz stereo audio output
- Official generation specifications
- 4–15 seconds, 24 FPS, base resolution of 768 pixels on the shorter side; actual local output depends on workflow settings
- Generation variant
- FL2VA: 0 images for text-to-video, 1 for image-to-video, 2 for first/last frames; Ref2VA: multimodal reference generation
- License
- MiniMax-H3 Community License
- Model files used here
- FL2VA / Ref2VA; 281 files, 498.475 GB total
This four-device validation includes video generation with audio from text. The timed runs below use reference inputs and do not constitute a complete text-to-video performance matrix. The official online service's 2K workflow does not establish the native output capability of this local setup.
Model information checked on 2026-09-20
02 / Deployment
MegaCube deployment configuration
- Compute devices
- 4 Cubes run TP4 inference
- Service interface
- The original setup hosts the shared ComfyUI interface on a separate device, which is excluded from the four inference nodes
- Runtime environment
- PyTorch 2.13.0+cu130；SGLang 0.5.18；ffmpeg
- Network
- Cross-node RoCE communication; dual NCCL interfaces and a Gloo management interface
- Application integration
- Asynchronous video API & ComfyUI
Prepare the compute nodes
Use matching software environments on all four devices and check the management network, RoCE, and ffmpeg.
Download and verify
Download MiniMax/MiniMax-H3 and verify all 281 files and SHA256 checksums.
Start the model variant
Synchronize FL2VA / Ref2VA weights and paths, then start the target variant with TP4.
Submit a generation task
Submit tasks, check progress, and save audio/video through the API or ComfyUI.
03 / Performance tests
MegaCube performance test results
2026-09-04; four inference nodes. Video generation with audio from text passed functional validation. The two timings below are Ref2VA runs using reference inputs; both produce 1344×768 video lasting approx. 5.17 seconds, with 50 sampling steps.
| Validation item | Method / conditions | Result |
|---|---|---|
| Weight integrity | 281 files, 498,474,765,023 bytes, full SHA-256 verification | Passed |
| T2VA | Text input, 768p, 50 steps, synchronized audio/video generation | Passed |
| FL2VA | First/last-frame input pipeline | Passed |
| Ref2VA | Reference image, reference video, and video audio-track encoding | Passed |
| Asynchronous video API | Submission, status polling, and content download | Passed |
| ComfyUI | Custom nodes, Save Video, user login, and asset / output isolation | Passed |
| Task | Video specifications | Sampling steps | Generation time |
|---|---|---|---|
| Task A | 1344×768 · Approx. 5.17 s | 50 | 16.28 min |
| Task B | 1344×768 · Approx. 5.17 s | 50 | 16.20 min |
- These two generation times are individual observations, not a mean or P95, and do not constitute a complete text-to-video performance matrix.
- The demo video lasts 10.125 seconds; the timed generation cases above produce 5.17-second videos.
Plan model deployment and application integration
Evaluate device configurations and inference services for your model size, workloads, and concurrency needs.