Self-hosting MiniMax H3

H3-Base is a 33B-parameter open-weights model. A 4× H200 node renders a 5-second 768p clip in about 75 seconds at 50 steps; the 8-step Turbo LoRA is roughly 5× faster with softer output. Rented nodes are billed per GPU-hour.

https://github.com/MiniMax-AI/MiniMax-H3

GPU presets used by the calculator

Node	5s clip @ 50 steps	Reference $/GPU-hour	Sources
4× H200	75s	$3.50	platform.minimax.io
4× H100	80s(estimate)	$2.50	github.com
8× B300	19s	$8.00	github.com
2× RTX 5090	560s	$0.70	platform.minimax.io

Clips per dayClip length (seconds)ResolutionMode

Swipe sideways to see all columns

Provider	Resolution	Per clip	Per day	Per month (30 days)
fal · H3 Max H3 Max offers text-to-video and image-to-video only; reference-to-video is not available yet.	480p	Not supported
fal · H3 MaxCheapest H3 Max offers text-to-video and image-to-video only; reference-to-video is not available yet. Native output is 1344×768 at 24 fps with synchronized audio.	768p	$0.10	$10.00 Standard: $40.00	$300.00 Promo until 2026-09-07
fal · MiniMax H3 First 5 reference images are free; each additional image costs $0.08.	480p	Not supported
fal · MiniMax H3 First 5 reference images are free; each additional image costs $0.08.	768p	$0.30	$30.00	$900.00
fal · MiniMax H3 First 5 reference images are free; each additional image costs $0.08.	2k	Not supported
fal · MiniMax H3 First 5 reference images are free; each additional image costs $0.08.	4k	Not supported
MiniMax official API 768p is listed at $0.08/s but is in closed beta; during beta the API bills at the 2K rate of $0.13/s, which is what this calculator uses. First 5 reference images are free; each additional image costs $0.04. Reference video input is billed by its duration at the output resolution rate.	768p	$0.65	$65.00	$1,950.00
MiniMax official API First 5 reference images are free; each additional image costs $0.04. Reference video input is billed by its duration at the output resolution rate.	2k	Not supported
WaveSpeed · MiniMax H3 Reference video input is billed at $0.13 per second; images beyond the first 5 cost $0.04 each.	2k	Not supported
Self-host H3-Base 4× H200 · 1 node · ≈75s per clip · capacity 806 clips/day per node	768p	$3.36	$336.00	$10,080.00

Self-host assumptions

GPU nodeGPUs per nodeUSD per GPU-hourUtilization (70%)SamplingBilling

Timings come from single-request SGLang benchmarks at 1344×768, 124 frames, 50 steps. Real throughput depends on batching and concurrency; treat all self-host numbers as estimates. 480p uses the same 768p timings (no speed-up assumed).

Break-even

Standard prices; promotions excluded.

At these assumptions fal · MiniMax H3 768p stays cheaper than self-hosting on 4× H200 at any volume.

What hardware do I need?

MiniMax's reference deployment is 8× B200. Documented alternatives include 4× H200, 4× H100 and 2× RTX 5090 with layerwise offload.

Can I generate 2K when self-hosting?

Not yet. The open release produces 768p; the 2K regenerate stage is API-only.

When is self-hosting cheaper?

Only when a node is kept busy. Use the calculator: with dedicated billing the break-even is the volume where one node's daily rent equals the hosted per-clip price times clips per day.