fal's H3 Max: Faster Hailuo, Capped at 768p
fal Research post-trained MiniMax H3 and shipped it on 27 August at $0.08/sec. It is faster and cheaper than base H3 — and it gives up 2K to get there.
MiniMax shipped H3 on 31 July 2026: a multimodal video model that takes text, images, video and audio in one context and returns up to 15 seconds of 2K with native stereo sound. On 27 August, fal Research announced H3 Max — its own post-trained version of that model, tuned for prompt adherence and aesthetics and co-optimised with fal's inference stack.
The framing everywhere is "faster and cheaper." That part is true. The part worth reading carefully is what it costs you to get there, because H3 Max is not a strictly better H3 — it is a different point on the same curve, and which end you want depends entirely on whether the output is a test or a deliverable.
What each one actually is
MiniMax H3 is the frontier model. Its distinguishing feature is not resolution but context: text, image, video and audio all go in together, which is what makes instruction-guided editing, brand and text rendering, and video-to-video motion transfer work in a single pass rather than as a chain of separate jobs. Up to 15 seconds, 2K, native stereo. It runs from about $0.13/second at source.
H3 Max is fal Research's post-train of that model. New training data for stronger prompt adherence and better aesthetics, then co-optimised against fal's own serving stack for speed. Output is 480p or 768p, with 768p the default — at 16:9 that is 1344×768 at 24fps, with native audio, from 5 to 15 seconds. It lists at $0.08/second at 768p, which is $4.80 a minute, with no subscription and no minimum.
The trade nobody puts in the headline
Base H3 does 2K. H3 Max caps at 768p.
That is a real drop, and it decides the use case for you more than the price does:
- 768p is fine for a vertical social cut, a hook you will overlay captions on, a client preview, or any of the twenty variations you generate before choosing one.
- 768p is not fine as a delivered master, a paid-ad hero at 4K, anything going to a large screen, or anything a client will grade.
The honest read: H3 Max is an iteration model that happens to be good enough to ship for social. Base H3 is a finishing model. Reaching for H3 Max because it is cheaper per second, then discovering you cannot deliver at the resolution the brief asked for, converts a saving into a re-render.
What the pricing actually means
fal launched H3 Max with 50% off for its first 14 days, which puts it near $0.04/second during the promo. Introductory pricing is not a planning number — build your budget on $0.08 and treat the discount as a reason to test now rather than a rate to forecast on.
There is also a genuinely useful free tier: five 5-second 768p generations a day with native audio without an account, and five more per day signed in, at up to 15 seconds, resetting on a rolling 24-hour basis. That is enough to answer "does this model hold my subject" before any money moves, which is the only question worth answering first.
| MiniMax H3 | H3 Max (fal) | |
|---|---|---|
| Shipped | 31 July 2026 | 27 August 2026 |
| Max resolution | 2K | 768p (480p option) |
| Duration | up to 15s | 5–15s |
| Audio | native stereo | native |
| Multimodal input | text, image, video, audio | post-trained on the same base |
| Rate | from ~$0.13/sec | $0.08/sec at 768p |
| Best for | finishing, editing, 2K delivery | iteration, social cuts, volume |
Using H3 on Versely
Versely runs MiniMax H3 text-to-video at 26 credits, generating at 2K — the base model rather than the 768p post-train, because the jobs it gets pointed at are usually the ones where resolution is the deliverable.
The workflow that makes the resolution question moot: iterate cheap, finish expensive. Generate your variations wherever a 768p test frame answers the question, decide which subject and motion actually work, and only then spend a 2K render on the one that is going out. That is the same discipline as locking the still before you spend on motion — the cost that hurts is never the first render, it is the fifth one at full resolution.
If you are choosing between models for a specific shot rather than in the abstract, the AI video generator exposes the H3 family alongside the rest of the catalogue, so the comparison is per-job rather than per-subscription.
Why a post-train is interesting at all
H3 shipped with open weights, and H3 Max is the first widely-visible demonstration of what that unlocks: an inference provider taking a frontier model, training on top of it, tuning it against their own serving stack, and shipping the result as a distinct product at a different price point.
That is a different competitive shape from the closed-API era, where a model was a fixed thing you called. It means the interesting question stops being "which lab is ahead" and becomes "whose post-train of which base model fits this job" — and it means a model's published spec is now a starting point that someone else can move.
Expect more of these. It also means the version string matters: "H3" and "H3 Max" are not tiers of one product, they are two models with different ceilings.
FAQ
Is H3 Max better than MiniMax H3? Not universally. It is faster, cheaper per second, and post-trained for stronger prompt adherence — but it caps at 768p where base H3 does 2K. For iteration and social output it is the better pick; for anything delivered at high resolution it is not.
How much does H3 Max cost? $0.08 per second at 768p, or $4.80 a minute, with no subscription or minimum. It launched with 50% off for 14 days, so early testing ran nearer $0.04/second — build budgets on the list rate, not the promo.
Can I try H3 Max for free? Yes. Five 5-second 768p generations a day with native audio without an account, plus five more per day signed in at up to 15 seconds, on a rolling 24-hour reset.
What resolution does H3 Max output? 480p or 768p, with 768p the default. At 16:9 that is 1344×768 at 24fps with native audio.
What makes MiniMax H3 different from other video models? It takes text, images, video and audio in one context, so instruction-guided edits, brand and text rendering, and video-to-video motion transfer happen in a single pass instead of a chain of separate jobs.
The takeaway
H3 Max is a good model at a good price with a ceiling you have to plan around. Use it where 768p is genuinely enough — which is most of what gets posted — and keep base H3 for the render that has to survive a client, a grade or a large screen.
The version names invite you to read Max as the upgrade. On resolution, it is the opposite.