We’re open-sourcing Native Video Inference: the Rust + CUDA runtime we built at
@Creatify_AI to serve video models.
Cold start to first denoising step: ~7s on a fresh B200, vs 65–75s on
@sgl_project Diffusion and vLLM-Omni. Same output, bit for bit.
No PyTorch at inference. Kernels compiled ahead of time. Everything loads in parallel.
On our production traffic, that means 84% less idle GPU time.
Supports MiniMax H3, Wan 2.1 and LTX 2.5. Code and write-up below 👇