H3SERVE

MiniMax-H3, 8-step LightX2V Turbo LoRA, first-frame-conditioned (fl2v) video generation. This is an async job endpoint, not a realtime stream — the measured ledger puts one clip at ~33s of GPU compute (7.7x over the 5.17s of video it produces) on today’s best single-card config. Submit a prompt, watch the honest progress below, and the finished clip plays automatically.

No image selected — the server falls back to its own bundled reference image.