LingBot-Video's 8-step DMD (Distribution Matching Distillation, a few-step generation distillation method) is still a 30B-parameter MoE (Mixture of Experts, where only a subset of parameters activates per inference): fewer steps doesn't mean less VRAM.Our cross-check of the official tables shows FP8 (low-precision format) shrinks single-inference time from 46.76s to 36.72s, yet per-GPU VRAM rises from 83.75 to 110.80 GiB.
What this is
A3B indicates roughly 3B parameters activate per inference; it does not mean only 3B weights are stored. 8 steps only reduces sampling iterations—you cannot infer a 5x speedup from that. The official DMD also requires specially trained student weights (parameters obtained through distillation) and matched scheduler settings; simply switching a regular model to 8 steps is not correct usage.
Industry view
We believe what's worth borrowing is not the "8-step" label but the practice of recording model identity, sampling settings, and hardware layout separately; deriving VRAM and cost directly from step count risks choosing wrong from the very first experiment.
The risk: this article only cross-checks source code and documentation, with no A6000 testing done; a single official test cannot prove that all FP8 paths consume more VRAM, and speed and quality at other resolutions remain unverified.
Impact on regular people
For enterprise IT: When evaluating video models, weigh load time, peak VRAM, and end-to-end cost together—fewer steps does not equal cheaper.
For working professionals: Fewer-step generation may shorten waits, but cloud billing and quality still depend on model configuration; self-deployment still carries hardware thresholds.
For consumer markets: Faster generation does not yet mean cheaper or clearer video—whether added supply translates into price drops remains to be seen.