License and acknowledgements
Vflash source code
Vflash is released under Apache License 2.0. The source repository contains no model weights.
Models and adapters
MiniMax H3 and optional LoRA checkpoints retain their own licenses, usage rules, and geographic terms. The Vflash code license does not grant rights to those weights or override their restrictions. Read the terms for the exact model and adapter revision you use.
Thanks
Vflash builds on the published H3 architecture and checkpoints from MiniMax. LightX2V's open inference work and Turbo adapters provide a valuable compatibility source and comparison reference.
The runtime uses PyTorch and Triton, with NVIDIA CUDA. We also acknowledge the work of NVIDIA Sol-Engine, FlashAttention, CUTLASS, SageAttention, FastVideo, Alibaba PAI, and OpenVDN that informed H3 acceleration research. Vflash's direct two-GPU relayout uses the destination-major technique described by Sol-Engine, adapted to Vflash's own collective layout. The alignment page records what was adopted, deferred or rejected.
Acknowledgement does not imply that every referenced implementation or adapter is included in this release. The supported runtime configurations are listed under profiles and hardware.
The complete pipeline uses Diffusers, Transformers, Accelerate, PEFT, and the official decoder code distributed with the pinned MiniMax H3 model. Local MP4 delivery uses FFmpeg. Each dependency retains its own license; the Vflash repository does not redistribute their model payloads.
Complete pipeline dependencies
The complete image includes the pinned Diffusers, Transformers, Accelerate and PEFT adapters. Their licenses and package notices remain in the installed distributions.
MP4 encoding uses the Debian FFmpeg package. Its copyright and license notice is included at /usr/share/doc/ffmpeg/copyright in the image; corresponding source packages are available from Debian Sources. These dependencies retain their own terms. Official H3 decoder code is loaded from the separately verified model snapshot.