fal H3 Max Turbo vs H3 Max: What We Measured

This page is for builders choosing an endpoint for batch rendering or an interactive channel. It does not repeat what a post-trained hosted model is or how to write the first API call — those live on the H3 Max reference page, along with the JavaScript quickstart. What follows is only what this site could time, count or read off a card.

Three things fall out of that table. First, the gap between Turbo and Max is smaller than the price gap: 1.10 seconds of inference against 1.75, but 4.3 seconds of wall time against 6.0, because a fixed overhead of roughly 2.5 to 3 seconds sits on top of both and does not scale with clip length. Second, the 95th percentile on every balanced row is longer than the clip it produced, which is the number that decides a player's buffer. Third, fal's published card samples run slower than what arrived here — 1.55 against 1.10 on Turbo, 2.53 against 1.75 on Max — so the card is not an optimistic figure to be discounted; it is a different measurement of a different job.

Buffer rule this supports: under balanced expansion, start rendering two clips ahead rather than one, and consider making the second of those two a disabled job, whose 95th percentile of 4.2 seconds leaves real headroom inside a 10-second slot. Longer clips amortise the fixed overhead better: at 480p the same endpoint gave RTF 0.85 for a 5-second clip, 0.53 at 8 seconds and 0.43 at 10.

Why the local column is not a speed comparison. The RTX 3060 run rendered 124 frames at 1344×768 with the four-step adapter in 526.7, 528.9 and 528.6 seconds, against 2,175.5 and 2,176.2 seconds with the adapter off at twenty steps. Those are wall times for a different canvas, a different frame count and a different machine than fal's backend, so this page states them beside the hosted figures and never divides one by the other. The full conditions sit on the Turbo LoRA page.

What turning it off costs you. The saving is not free, and the price is not money — fal bills output seconds, so a disabled job and a balanced job of the same length cost exactly the same. What changes is the picture. Without the rewriter the clips arrive as a single continuous take with the look of a CG toy render, where balanced clips read as cinematic live action, because the rewriter put "Cinematic" into all 165 expanded prompts inspected and "live-action" into 113 of them. The automatic soundscape and music paragraphs disappear too. And you cannot get the best of both by pre-expanding: feeding an already-expanded storyboard back in under balanced does not skip the rewrite, it rewrites the rewrite, growing a 1,940-character input to between 2,356 and 2,525. Only disabled actually skips it, which makes "write your own shot list, then send it with expansion off" the low-latency way to keep a rich prompt.

The rewriter is also your concurrency ceiling. An earlier run had concluded that one fal key gets about three useful lanes, because pushing concurrency to 10 sent queue time to a 72-second 95th percentile. Two identical batches twenty seconds apart settled where that serialisation lives: at concurrency 10 with balanced, overhead ran 8.4 seconds at the median and 15.5 at the 95th percentile; with disabled, the same ten parallel jobs held 0.9 seconds flat. Inference stayed at 1.07 to 1.09 seconds on both sides. It is the rewriter that serialises, not the video backend — with it off, thirty 10-second clips finished as a batch in 23.6 seconds with no queueing at all. Plan lanes accordingly: about three under balanced, at least ten under disabled, and nothing above ten has been tested here.

Seeds do not bring a clip back. Re-submitting byte-identical inputs at the same seed produced structural similarity of 0.49 to 0.56 and peak signal-to-noise around 15.5 to 16.7 dB against the originals — different footage, not a variation. With expansion disabled, where the rewriter's randomness is gone, two same-seed passes still landed at 0.48 to 0.79. So the seed is a draw, not an address. Cache the mp4 file itself; a prerender cache, a branch replay or an A/B comparison keyed on seed will not find its clip again.

Across 69 same-seed clips of one described character, 19 kept the shape the prompt asked for. Splitting by whether the prompt itself carried an identity cue explains all of it: clips whose text mentioned the single glowing eye held 14 of 15, clips without any such cue held 0 of 45. The drift pattern was the same on Turbo, on Max and at 768p, so it is not something Turbo's distillation introduced. Turning the rewriter off does not rescue it either — 30 disabled clips of the same story held 5, no better than balanced, and the cued scenes actually did worse. The model's prior for "a brass robot" is a two-eyed humanoid, and nothing in the prompt pipeline defends against that except the prompt.

The full ladder — including how to write shot lists that survive disabled mode, and how to size a buffer around each option — belongs to this site's interactive-channel tutorial, which is being written and is not published yet. This section deliberately stops at the choice between the three approaches above.

No video is embedded on this page and none is hosted here. The clip packs behind these strips are published as a GitHub release, and the strips themselves were cut from those files with ffmpeg at fixed timestamps — no colour correction, no crop, no re-ordering.

Where it is: 03-dev/tools/fal-batch in this site's repository, with the prompt lists used for the runs above. The clips it produced are attached to the clips-2026-09 release. The estimate is arithmetic and the dashboard is the invoice; on the batch behind this page the two agreed exactly.

Conditions attached to every timing on this page. One fal account, two consecutive nights (3 September around 21:00 and 4 September around 01:00, US Pacific), one region, the queue API at queue.fal.run with a 1.5-second poll interval, concurrency 3 except where stated. That poll interval puts up to 1.5 seconds of quantisation error into the queue and overhead columns. Group sizes run from 3 to 30, no statistical significance is claimed, no picture-quality judgement was made, and downloads are excluded from wall time. During one stage the account balance ran out and fal returned intermittent 403s; those submissions were not billed and are excluded from every figure. Independent guide; not affiliated with MiniMax or fal.