CheaperGen

Accelerated inference

Same model. Minutes become seconds.

MiniMax H3 Fast runs MiniMax H3's open weights on our own GPU inference for open-source models. Below, both generate the same 8 prompts with the same settings, through our API — played in sync, with how long each took and what it costs.

Watch 7.6× fasterMiniMax H3 Fast API
median speed-up
7.6×
same prompt, same settings
median time to a finished clip
17s
vs 2m 08s for MiniMax H3
per 6-second clip
$0.328
vs $0.342 (−4%)
prompts compared
8
motion, people, text, physics

Pick per job

Both stay available. You choose.

The original for MiniMax's own output and 2K; the accelerated build for drafts, iteration and volume. Each has its own model id, price and page.

OriginalMiniMax H3

Hailuo H3 — expressive motion and physics, up to 2K

from$0.057/sOfficial price $0.08≈ 2m 08s per clip
MiniMax H3 API Accelerated variantMiniMax H3 Fast

MiniMax H3 in seconds, not minutes — an accelerated build of the open weights

from$0.037/sOfficial MiniMax H3 API $0.05vs. official MiniMax H3 API≈ 17s per clip
MiniMax H3 Fast API
Accelerated variant

An inference-accelerated build of MiniMax H3's open weights, served on our own GPU inference for open-source models. The acceleration is not a MiniMax release, and MiniMax does not sell this variant.

The original: MiniMax H3 — choose it when you need MiniMax's own output or its higher tiers.

How it is faster

  • Same H3 open weights and inputs: text, first / last frame, or references
  • Low-precision attention and kernels, plus post-training for fewer, faster sampling steps (exact recipe not published)
  • Served warm on GPUs that already hold the model, and returned in the same request
  • Up to 768P (no 2K) and 4–15 s; output can differ from MiniMax's own H3 API for the same prompt

Built on MiniMax H3's open weights, released by MiniMax. Use is subject to MiniMax's usage policy and our Terms of Service.

How we list accelerated models

Faster, and labelled as such.

  • Its own model

    An accelerated build gets its own name, id, page and price. It never stands in for the original.

  • Never swapped in

    Routing happens within a model, so a request to the original only ever runs on routes that serve the original.

  • Shown, not claimed

    Speed and price come from real generations through our API, side by side with the original, with the method written down.

FAQ

Questions, answered.

  • One API for the original and the accelerated build
  • Pay per second, failed generations refunded
  • Prices shown before you run
Try MiniMax H3 FastHow our inference works

An open-weights model served on our own GPU inference for open-source models — low-precision attention and kernels, fewer sampling steps and warm GPUs — so it returns the same kind of output much faster and for less. It is listed as its own model, next to the original.

It starts from the same open weights and takes the same inputs, but shortened sampling changes details, and some tiers (such as 2K) are not available. That is why we show every comparison side by side, from the same prompt and settings.

No. Routing happens within a model: a request to the original model only goes to routes that serve the original. You choose the accelerated variant explicitly, by its own model id.

Through CheaperGen's own API routes on the date shown, both sides submitted at the same moment with the same prompt and settings. The times run from submitting the request to a downloadable clip. Clips are shown as delivered, without re-encoding.