Runway Characters: serving a 57B-parameter video model in real time
Runway published research on serving its roughly 57B-parameter video diffusion transformer in real time, which now runs each session on four H100 GPUs.
By GennnAi··1 min read
The model must generate a four-frame 704x1088 chunk every 166.7ms for one user. Launch required eight H100s per session; optimizations cut the diffusion step from 240.9ms to 159.3ms on four H100s, and a fused kernel reaches 147.4ms. Cache upkeep fell from 69.7ms to 6.2ms per step. Lip-sync differences stayed within run-to-run variation, and the worker averaged above 23.5 fps in 99.5% of sessions over seven weeks.
Source: Runway, published on November 5, 2026.