FastH3 8-Step V2 Is Real — Its RTX 5080 Status Is Unverified

FastVideo (Hao AI Lab @ UCSD) released FastH3 8-Step V2, an 8-transformer-forward MiniMax-H3 distillation using data-free DMD2 and VSA-H3 sparse attention. We reviewed the official FastVideo model card and the current official ComfyUI documentation in detail.

The original FastVideo repository reports roughly 35B BF16 parameters, which gives a roughly 70GB weight-size calculation. That calculation is not an official minimum-VRAM requirement or a NeuralDrift measurement. Current official ComfyUI documentation provides native FastH3 templates and a separately named pruned INT8 diffusion model, but does not state RTX 5080 compatibility or required VRAM. NeuralDrift has not attempted a run, so its local fit, generation time, and A/V sync remain unverified. Full sourcing and the GPU compatibility matrix are on the Lab writeup below.

Current official ComfyUI documentation lists FastH3 text-to-video and image-to-video templates with optional first/last-frame conditioning; it says Ref2VA was not distilled. Base MiniMax-H3 also supports reference-to-video and has its own native ComfyUI templates. These are upstream capability statements, not NeuralDrift execution results.