Tools · Model Watch
Runway shows what it takes to keep a generated character talking in real time
The company describes the GPU and latency engineering behind a 57-billion-parameter conversational video model.
Reporting updated Nov 11, 2026

THE TAKEAWAY
A live video character cannot pause for a render queue: it must make the next frame while the conversation keeps moving.
WHAT HAPPENED
Runway’s November 5 engineering report details how its approximately 57B-parameter Characters model generates 704×1088 video around 24fps. By reducing repeated cache processing and overlapping communication with computation, it reports a reduction from 240.9ms to 159.3ms for its relevant four-frame generation step on four H100 GPUs; a fused path reaches 147.4ms. These are first-party measurements, accompanied by caveats about quality evaluation and production rollbacks.
WHY IT MATTERS
Instead of comparing only visual demos, creators can understand why continuously responsive generated performers are technically difficult and expensive. It also clarifies why a usable character experience depends on audio turn-taking and inference costs.
Reader card
Four-frame generation deadline
cache reuse
communication/kernel fusion
measured serving and failure cases
VEYR ANGLE
Keep this engineering-led, with comparisons carefully attributed to Runway rather than presented as independent benchmarks.

