0.22.1Current release
MiniMax H3 video, natively.
From text or images to 4 MP with sound.
A native PyTorch/Triton denoising runtime that turns one request into a 5–15 second MP4 with audio: text, first frame, last frame, both, or up to nine reference images, up to 2720×1536 when image-conditioned. Python and the CLI deliver MP4; native HTTP returns latents. A separate 96 GiB SM120 API combines 4+6 visual sampling with twelve-step Larry native audio.
- Execution
- One GPU · resident or block ring
- Measured paths
- SM89 profiles; hybrid v0.1 up to 15 s and native 4 MP
The reference platform. BF16 hybrid Veda 44/6 measured about 660 s for 2720×1536 at five seconds and about 112 s for 928×512 with six-step audio refinement.
vflash plan i2va-turbo4-v01-544-exact-sm89 \
--gpu 0What you can generate
T2VA, Ref2VA with 1–9 reference images, I2VA, L2VA, FL2VA and short reference videos. Five to fifteen seconds at 24 fps; canvases are multiples of 32, up to 1 MP for ordinary requests and up to 2048² pixels with a 2720-pixel long edge for image-conditioned native HD, within the joint width × height × frames ≤ 2048² × 124 budget. Each profile and GPU qualifies its own subset. See modes and limits
Start with what you need
Install and check
Install the lightweight CLI and confirm your GPU and profile with doctor and plan. No weights needed.
02Generate your first video
Prepare the official model assets and create an MP4 with sound from text and images via CLI, Python or Docker.
03Segmented video and native audio
An explicit Python runtime on complete 96 GiB SM120: Veda, four Larry plus six original visual evaluations and twelve Larry audio evaluations.
Understand it. Build on it.
The native denoiser, LoRA merging, memory scheduling, two-GPU layouts and media delivery are Vflash code; encoders and VAEs use pinned official implementations.