Self-host an open-weights video model or pay per second?

LTX-2.5 is reported to have open weights plus an API, and MiniMax H3 weights are reported out since Aug 3, while its Context-IR and Regenerate-2K features stay hosted (reported by Magic Hour, read 2026-10-07). Neither LTX nor Hunyuan is in the Sume catalog, and Wan 3.0 on Sume is the hosted id. So the honest comparison is hosted Sume against your own GPU.

The formula

Let g be your GPU cost per hour (your number, not a vendor claim), t the GPU seconds one clip needs, and s the clip seconds. Self-hosting costs g x t / 3600 per clip, before storage, queueing, idle time and engineering. Sume's Wan 720p costs $0.125 x s, so the GPU-hour budget per clip-second is:

def breakeven_gpu_seconds_per_clip_second(gpu_dollars_per_hour,
                                         sume_rate=0.125):
    # GPU seconds you may spend per output second
    return sume_rate / gpu_dollars_per_hour * 3600
for g in (2.0, 4.0, 8.0):
    print(g, round(breakeven_gpu_seconds_per_clip_second(g), 1))

Read it

With g = $2 per hour the budget is 225 GPU seconds per output second; at $4 it is 112.5; at $8 it is 56.3 (Sume 720p rate, read 2026-10-07). If your measured render time per output second is below that and the GPU stays busy, self-hosting wins on compute alone. Idle hours push the other way.

• Utilisation matters more than the sticker rate: a GPU busy 20 percent of the day costs five times the quoted rate per busy hour.
• Hosted-only features (Context-IR, Regenerate-2K) are not in the weights.
• Sume bills per job with a reserve and refund on failure, so a failed render costs nothing.