Backend inference
timings.inference is the value returned with an eligible fal result. The live OpenAPI schema defines it as DiT denoising time on the GPU backend. It can be absent on routes that do not report backend timings.
SOURCE CHECKED · SEP 7, 2026
MiniMax H3 Max
Generation Time
fal reports a 5-second 768P video in under 3 seconds of wall time. That is faster than playback, but it is not a promise that every end-to-end request finishes in three seconds.
PUBLISHED WALL TIME
<3.0s
5-second · 768P video
Faster than real-time playback under fal's published test condition.
01 / READ THE RIGHT CLOCK
Three numbers can all be called “generation time.”
The MiniMax H3 Max generation time only becomes useful when the measurement boundary is attached.
Published wall time
fal's launch article says a 5-second video completes in under 3 seconds of wall time. It does not publish a service percentile, queue-inclusive guarantee, or SLA around that statement.
End-to-end wait
This is the interval a user feels: submit the request, wait through queueing and any prompt rewrite, receive the response, then make the media usable. It varies with the route, settings, traffic, network, and client.
02 / WHAT IS ACTUALLY PUBLISHED
Known numbers. Visible gaps.
This MiniMax H3 Max generation time ledger does not extend one benchmark into durations fal has not benchmarked the same way.
Output
Published evidence
Do not assume
5 seconds · 768P
Under 3 seconds of wall time in fal's launch claim; one official sample reports 2.53 seconds in
timings.inference.
Neither number guarantees the complete queue-to-download wait.
5 seconds · 480P
Supported by the endpoint.
Not published: a separate fixed 480P latency or guaranteed speedup.
10 seconds
Supported within the 5–15 second duration range.
Not published: a formal 10-second wall-time benchmark or SLA.
15 seconds
Supported by the endpoint; fal's current model page describes roughly 15 seconds for this clip length.
Not an API guarantee: the schema sets duration limits, not a latency contract.
The endpoint accepts 5–15 second outputs at 480P or 768P. Video length is an input setting; generation time is an observed result. MiniMax H3 Max generation time cannot be derived from video length alone.
03 / ONE OFFICIAL RESULT
2.53 seconds is a sample, not an average.
The response displayed on fal's text-to-video endpoint pairs this 5.18-second, 1344×768 output with an inference value of 2.5285604159580544 seconds. Rounded for reading, that is 2.53 seconds.
It is useful because the output, settings, media, and backend timing appear together. It still represents one request. This site did not submit it, time it independently, or observe its queue and download path.
Inspect the official sample and response ↗04 / BEFORE THE RENDER
Prompt expansion can be longer than inference.
The OpenAPI description says balanced expansion returns in about a second, while quality mode can spend up to roughly 30 seconds preparing a richer prompt.
Balancedabout 1sschema description
Qualityup to ~30sschema description
Queue + transfervariablemeasure your route
These are separate stages, not numbers to add into a universal total. A provider can change queue behavior, preprocessing, hardware, or delivery without changing the model name.
05 / MEASURE YOUR REQUEST
Record both clocks.
To measure MiniMax H3 Max generation time for your own workflow, keep the request and the measurement boundary together.
- 01
Start before submission.
Record a monotonic client timestamp immediately before the request is sent. Save the endpoint, duration, resolution, input mode, and prompt expansion setting.
- 02
Keep the backend timing.
When the result returns, store
timings.inferenceif the route provides it. Do not rename that field “total time.” - 03
Stop when the output is usable.
Record the second timestamp when the response and playable media are available to your application. The difference is the end-to-end wait for that run.
- 04
Compare like with like.
Change one variable at a time. A different duration, resolution, route, prompt-expansion mode, account load, or network path is a different timing condition.
06 / SOURCE LEDGER
Every number has a boundary.
Launch benchmark5s video · under 3s wall timefal · checked Sep 7, 2026 ↗
Model landing pageOfficial sample · timing contextfal · checked Sep 7, 2026 ↗
Live OpenAPI schematimings.inference · prompt expansionfal · checked Sep 7, 2026 ↗
Official guideDuration · resolution · workflowfal · checked Sep 7, 2026 ↗
07 / FAQ
Timing, without the shortcut.
What is the published MiniMax H3 Max generation time?
fal reports that H3 Max can generate a 5-second 768P video in under 3 seconds of wall time. The claim is tied to that output length and resolution; it is not an end-to-end SLA for every request.
Is the 2.53-second sample timing the full request time?
No. The official sample response reports 2.53 seconds in timings.inference. fal's OpenAPI schema defines that field as DiT denoising time on the GPU backend, so it does not by itself describe queueing, prompt expansion, network transfer, or media download time.
How long does a 10-second H3 Max video take to generate?
fal's endpoint supports a 10-second output, but the launch benchmark does not publish a guaranteed 10-second generation time. Measure the route and settings you plan to use instead of multiplying the 5-second result.
How long does a 15-second H3 Max video take to generate?
The endpoint supports up to 15 seconds. fal's model landing page describes roughly 15 seconds for a 15-second clip, but the API schema does not make that an SLA or latency guarantee.
Does 480P always generate faster than 768P?
The endpoint supports both resolutions, but fal's headline timing is specifically for 768P. No separate public guarantee on this page supports a fixed 480P generation time.
Why can my H3 Max request take longer than the published time?
The published result is not the whole delivery path. Queueing, prompt expansion, request settings, network transfer, and downloading the finished media can add time around backend inference.
READ THE RESULT, THEN RUN YOUR OWN
Measure the route you will actually ship.
MiniMax H3 Max generation time is fast enough to change iteration, but one published number cannot replace your own end-to-end measurement.