0.6.12Stable release
Native H3 inference.
Built for your GPU.
A MiniMax H3 engine measured on SM86 and SM89. Run Turbo or Base16 keyframe generation through Python, containers or HTTP.
- Execution
- One GPU · resident/streamed
- Available profiles
- Turbo4 / Turbo8 / Base16
Serve text, references and keyframes; the complete pipeline reuses one GPU across stages.
vflash plan ref2va-turbo4-exact-sm89 \
--gpu 0The current public interface
Ref/T2 Turbo retains five seconds; Base16 and optional 544p keyframes accept five through ten. Single-SM89 Base16 defaults to Sol. Original v0.1 keyframes permit explicit Sol but default to dense. Media evidence is scoped by mode, not a blanket quality guarantee. Check the prerequisites
Start with what you need
Check your setup
Install the lightweight CLI and inspect your GPU, memory and profiles. No weights required.
02Generate a video
Prepare official model assets and generate an MP4 from text, images or a short clip with Python or Docker.
03Evaluate the results
Separate loading, repeated requests and video delivery. Check quality against your task.
Understand it. Build on it.
Native PyTorch and Triton execution, with owned LoRA, memory scheduling and two-GPU collective layouts.