Two numbers, two axes. The latency figure is clip length over wall clock: five seconds of output in about three seconds, roughly 1.7x faster than playback [3][14]. The throughput figure is about 35 times MiniMax's official H3 endpoint [4]. Those do not measure the same property. If the 35x were a per-request latency comparison on identical hardware, MiniMax's endpoint would need something near 105 seconds for the same five-second clip [15]. The more ordinary explanation is batch depth, GPU count and queue policy, and fal's announcement does not say which [4][5]. Throughput without a denominator is a claim about someone else's capacity plan.
For the multiple to reach your bill, several things have to hold: same GPU class, same resolution and frame count, same sampling step count, the audio track in the same state, and enough concurrent requests to fill whatever batch fal is filling [11]. Traffic that arrives one request at a time from a UI collects the three seconds and none of the 35x.
The gate is the better engineering here. fal's inference group changed precision, sampling and execution, and kept a change only when fal's evaluations showed output quality held up [7]. Every knob in that list buys speed with fidelity, so the gate is load-bearing rather than ceremonial. The evaluations were head-to-head human preference studies covering overall quality, prompt understanding and aesthetics [6]. That is self-grading, and unavoidably so: once fal Research adds training data to the base checkpoint [2], there is no public model with the same identity left to score against.
The commercial mechanism follows from the same fact. When two hosts expose the same weights, competition collapses onto latency, throughput, uptime and price [8]. Change the weights and design the runtime around them, and a broader GPU platform cannot reproduce the endpoint by loading the public checkpoint [9]. That is the wall fal is building against Replicate, Modal, RunPod and CoreWeave [10], and it is a real wall rather than a positioning statement. It also marks a move up the stack from hosting other labs' models to shaping them [17].
It is paid for twice. fal pays in per-model engineering: post-training plus a runtime pass, per base model, which does not amortise across a catalogue the way generic serving does [2][7]. The customer pays in portability. H3 Max exists as fal's text-to-video and image-to-video endpoints [11], and fal's platform also distributes MiniMax's own models [19], so the exit from H3 Max leads back to weights that were deliberately changed [2], with prompts tuned against the modified ones.
Scope matters on the ranking too. The first-place claim covers specific image-to-video categories, not the broader text-to-video leaderboard [12], and the source material cuts off before naming the benchmark in full [13]. The comparison against 12 leading video models was internal [5]. Two endpoints also means two integrations, and the placement only speaks to one of them [11][12].
Adoption therefore turns on a measurement fal cannot run for you: p95 latency at your own concurrency, on your own prompts, with audio in the configuration you ship [11]. The three-second figure is cheap to test in an afternoon. The 35x needs fal's hardware and MiniMax's rate limits, so it stays a claim about their workload until someone publishes the denominator.
