Fast-H3-Omni benchmarked: realtime video in ~12s vs H3's ~5min, at 4x the cost
tobowers · x · 2026-10-14
FastVideo and nuvalab released Fast-H3-Omni, claimed as the fastest, highest-quality open-source post-trained version of MiniMax's realtime H3 video model, previewing on the Reactor API until Oct 20, 2026. It supports first/last-frame conditioning, audio input with strong dialogue, and up to 3 reference images.
Developer tobowers ran a first-hand 23-shot comparison against H3 (Style Max, 480p):
- Speed: H3 median render 4m47s per shot (45s–8m10s); Fast-H3-Omni median 12.4s (10.1–16.7s) — hundreds of times faster
- Cost: full 23-shot run $0.81 on H3 vs $3.09 estimated on Fast-H3-Omni ($0.0125/session-second), i.e. 4x pricier but near-instant
- The writeup shares the full structured prompt format (subjectdefinitions, retentionanalysis) used to keep shot consistency, making the workflow reproducible