Sign inMake a clip free
New

MiniMax H3 Max is live — 5–15s clips, 768P, sound included

See what a clip costs

MiniMax H3 Turbo Is Three Different Things

The name is attached to three separate pieces of software, two of them shipping side by side on fal at a 2× price gap. Getting them mixed up is the difference between a clip that costs $0.04 a second and one that costs $0.08.
Work out which one you came for.

  • Sense 1 · Hosted

    fal's H3 Max Turbo

    A shipping fal endpoint family, minimax/h3-max-turbo — its own model, not a nickname. fal describes it as a post-trained variant of MiniMax H3 tuned for stronger prompt adherence and better aesthetics, co-optimised with its inference stack for throughput.
    $0.025/s at 480p and $0.04/s at 768p — half what MiniMax H3 Max costs. Text to video and image to video only; no reference endpoint.
    Both fal builds, all five endpoints and both rate cards are on one page here, with four clips we generated on this endpoint and the timings off the same run.

    Came from fal's model list? This is the one you were reading — and it is the default engine on this site.

    Run it here

  • Sense 2 · A label

    MiniMax H3 Max, under its launch label

    A different fal model — the build post-trained with MiniMax and released 26 August 2026, at $0.05/s 480p and $0.08/s 768p. Artificial Analysis carried it under the label MiniMax H3 Turbo (768p), and fal's launch page still says so.
    The board has since relabelled the row, so the name now survives mostly in citations of it.

    Came from a leaderboard or a benchmark thread? This is your model. It also runs here — it is the one with a reference-to-video endpoint.

    The board, read out

  • Sense 3 · Local

    A Turbo LoRA on the open-weights base

    A community label, not an official MiniMax or fal release. It means running the open-weights MiniMax H3 for roughly four to eight denoising steps instead of the usual twenty-five or more, using a step-distilled adapter or an acceleration framework — normally in ComfyUI, on your own GPU.
    It accelerates the base model. It is neither of the two above.

    Arrived from a ComfyUI workflow or a .safetensors filename? That is this one. The base model's weights are published on Hugging Face; neither post-trained build's are.

This site is an independent third party and is not affiliated with MiniMax or with fal. Every claim below names the page it was read on and the date it was read. Prices are fal's published rates read on 4 September 2026; a launch promotion currently discounts both Max models by 75% until 7 September 2026, and the figures above are the rates the promotion returns to.

The short answer

Is MiniMax H3 Turbo the Same as MiniMax H3 Max?

Ask fal, and no. minimax/h3-max-turbo and minimax/h3-max are two separate endpoint families on two separate model pages. Read on 4 September 2026, Turbo is $0.025/s at 480p and $0.04/s at 768p; Max is $0.05 and $0.08 — an exact 2× gap at both resolutions. Turbo publishes no reference-to-video endpoint; Max does.

Ask the leaderboard, and yes — but about the other model. fal's launch page states that Artificial Analysis listed MiniMax H3 Max "under its internal name, MiniMax H3 Turbo (768p)". The same two words are at once a live product name and a retired label for the product beside it.
Opened on 3 September 2026 the board reads Minimax H3 Max (post-trained by fal), so a reader following fal's sentence will not find the word Turbo there any more. Nobody has written that down, which is why this page exists.

In a local generation thread, neither. There the word means step-distilled LoRAs that run the open-weights base model in a handful of steps on your own GPU. Different artefact, different hardware, different output.
The short version: run the hosted ones, or read the base model's card.

Elo · first with audio
1,201

Elo · first with audio

Resolution · no 2K
480–768P

Resolution · no 2K

Duration · 24 fps
5–15 s

Duration · 24 fps

Things called Turbo
3

Things called Turbo

  • No 2K on either Max build
  • No published Max weights
  • No 4-second clips

Every name in circulation

Which MiniMax H3 Name Refers to What

Twelve strings, five things. This is the table nobody publishes, and the reason a search for one of these names so often lands on a page about a different model.

Rule of thumb: fal now sells two post-trains side by side — H3 Max Turbo and MiniMax H3 Max — and Turbo is the cheaper, newer one, not a nickname for the other. The word also survives as the leaderboard's launch label for MiniMax H3 Max, and as a community name for step-distilled LoRAs on the base model. Everything else here is an API identifier for one of those.

  • H3 Max Turbo

    What it actually is
    A hosted model of its own. $0.025/s 480p, $0.04/s 768p. No reference endpoint.Sense 1
    Where you meet it
    fal · here, as the default
    Run it

  • MiniMax H3 Max

    What it actually is
    A different hosted model. $0.05/s 480p, $0.08/s 768p. Has reference-to-video.Sense 2
    Where you meet it
    MiniMax docs · fal · here
    Run it

  • MiniMax H3 Turbo (768p)

    What it actually is
    Not the row above it. The label the board used for MiniMax H3 Max at launch.Sense 2
    Where you meet it
    fal's page still cites it

  • Minimax H3 Max (post-trained by fal)

    What it actually is
    That same board row, as it is labelled todaySense 2
    Where you meet it
    Artificial Analysis, 3 Sep 2026

  • MiniMax H3 fast · fast H3

    What it actually is
    Not an official name. People reaching for the speed build.
    Where you meet it
    Forums, chat, half-remembered

  • Turbo LoRA · step-distilled H3

    What it actually is
    Local acceleration of the open-weights base model, four to eight stepsSense 3
    Where you meet it
    ComfyUI workflows

  • MiniMax-H3-Max

    What it actually is
    API model id on MiniMax's own v2 endpoint
    Where you meet it
    platform.minimax.io

  • minimax/h3-max-turbo/*

    What it actually is
    fal's endpoint family for Turbo: text to video, image to video. Nothing else answers.
    Where you meet it
    fal.ai

  • minimax/h3-max/*

    What it actually is
    fal's other endpoint family: text to video, image to video, reference to video
    Where you meet it
    fal.ai

  • minimax:h3@max

    What it actually is
    Runware's model id for the same hosted model
    Where you meet it
    runware.ai

  • MiniMax H3

    What it actually is
    The open-weights base model Max was post-trained from. Does 2K, and takes video in.
    Where you meet it
    Hugging Face · MiniMaxAI
    Where we run it

  • MiniMax M3

    What it actually is
    Not a video model. MiniMax's ~428B language and agent model.
    Where you meet it
    A different product line

What the Max builds cannot do

  • No 2K, on either of them. 768P is the ceiling. 2K is a base MiniMax H3 capability, so that is what video to video runs here.
  • No four-second clips. Five seconds is the floor. The base model accepts four, which is exactly why people hit this.
  • No published weights. The base model's are open. Neither post-trained build's have been released.
  • No reference input on Turbo at all. fal publishes a reference endpoint for MiniMax H3 Max but not for H3 Max Turbo, and MiniMax's own API does not offer reference-to-video on the Max line yet. On this site, reference-to-video runs on MiniMax H3.

Speed is what comes back in exchange. fal reports a five-second 768P clip returning in under three seconds on its own infrastructure — its figure, on its hardware, not a measurement of ours.

Sources and method

Every label above was read on the page named beside it — fal's on 4 September 2026, the rest on 3 September. The first two rows are two different models one word apart at a 2× price gap; rows three and four are one row of one board, three weeks apart. That is the whole confusion, in four lines.

Read on

Third-party blind tests

The Board That Started the Turbo Confusion

Artificial Analysis runs blind head-to-head preference votes with public sample counts. This is the image-to-video board, with audio, opened on 3 September 2026.

  • Image to video · with audio

    First place

    #1

    9001500
    Elo
    1,201
    95% CI
    −9 / +9
    Samples
    5,469
    Board
    With audio
    Listed as “Minimax H3 Max (post-trained by fal)”
    Open the board

  • Design Arena · image to video

    #1

    9001500
    Elo
    1,341
    Base H3
    1,333
    As reported by fal, 25 August 2026 — not re-read by us

Read the rest of the column before quoting the win. Dreamina Seedance 2.0 720p sits eleven points behind at 1,190 with a tighter interval, and MiniMax H3 — the open-weights base this model was post-trained from — is third at 1,186. At ±9 the top three overlap.
The claim that survives scrutiny is narrower and more useful: it is first, and it is by a distance the fastest thing near the top. Not that it is decisively better looking.

The label is the part that moved. When fal published its launch page this row read MiniMax H3 Turbo (768p) — and fal's page still says so, dated 25 August 2026. Note which model that sentence is about: it is this row, MiniMax H3 Max, and not the separate H3 Max Turbo endpoint fal now also sells.

Artificial Analysis · image to video, with audio · opened 3 September 2026

How these boards work, and what ±9 meansMethod

Elo here comes from blind pairwise votes: two clips generated from one prompt, no model names shown, a human picks one. The confidence interval is the part that gets dropped when a rank is quoted. Two models whose intervals overlap have not been separated by the votes cast so far, however far apart their positions look in a numbered list — and a board with 5,469 samples on the top row and 17,728 on the second is not measuring both with equal precision.

The hosted models

What the Max Builds Actually Give You

Senses 1 and 2, which share every specification below except price and reference input. From MiniMax's own video generation guide and fal's endpoint schemas, read 3–4 September 2026.

Resolution
480P or 768P · 768P is the default and what it is tuned around · no 2K
MiniMax video generation guide
768P at 16:9
1344×768 at 24 fps
MiniMax video generation guide
Duration
5–15 s, whole seconds · four seconds is rejected
MiniMax video generation guide
Frame shapes
21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16 on text to video; image to video follows your picture
MiniMax video generation guide
Audio
Generated in the same pass as the picture, not dubbed on afterwards
MiniMax video generation guide
Prompt
Up to 7,000 characters
MiniMax video generation guide
Price, fal list
H3 Max Turbo $0.025/s at 480p, $0.04/s at 768p. MiniMax H3 Max $0.05 and $0.08. fal's published rates, read 4 September 2026
MiniMax video generation guide
Reference input
MiniMax H3 Max has a reference endpoint on fal. H3 Max Turbo has none.
MiniMax video generation guide
Weights
Base MiniMax H3 is open weights. Neither post-trained Max build has been published.
MiniMax video generation guide

Two of these are the usual trip-ups. Four seconds is not a valid length here — the base model accepts it and these do not, which is what catches people moving across. And there is no 2K at all: if you need it, the base model is the route, which is what video to video and reference to video run on this site.
A third is newer: a launch promotion is taking 75% off both Max models until 7 September 2026, so a price read off fal today is not the price the table above quotes. The figures here are the standing rates the promotion returns to.
What the fast build is actually for is the thing it made possible in the last week of August: channels that generate faster than they play. There are four of them and one of ours is playing here.

Questions people actually ask

MiniMax H3 Turbo FAQ

Which model the name means, what it can do, and where the weights are

Is MiniMax H3 Turbo the same as MiniMax H3 Max?

It depends which Turbo you mean, and there are three. fal ships H3 Max Turbo as its own endpoint family, minimax/h3-max-turbo, separate from minimax/h3-max and priced at half of it: read on 4 September 2026, fal lists H3 Max Turbo at $0.025 per second at 480p and $0.04 at 768p against $0.05 and $0.08 for MiniMax H3 Max. Those are two different models, so for that sense the answer is no. Separately, Artificial Analysis carried MiniMax H3 Max itself under the label MiniMax H3 Turbo (768p) at launch, and for that sense the answer is yes — one model, an older display label. And in a local generation thread Turbo means step-distilled LoRAs that accelerate the open-weights base model on your own GPU, which is a third thing again.

Should I use H3 Max Turbo or MiniMax H3 Max on fal?

Read on 4 September 2026, fal describes H3 Max Turbo as a post-trained variant of MiniMax H3 tuned for stronger prompt adherence and better aesthetics, co-optimised with its own inference stack for higher throughput, and prices it at half of MiniMax H3 Max at both resolutions. The one thing MiniMax H3 Max still has that Turbo does not is a reference-to-video endpoint: fal publishes text-to-video and image-to-video for Turbo, and text-to-video, image-to-video and reference-to-video for MiniMax H3 Max. Both cover 480p and 768p, 5 to 15 seconds, 24 fps, with audio generated in the same pass. Both descriptions and both prices are fal's own rather than measurements of ours.

Why does fal's page mention a name the leaderboard no longer shows?

Because the board relabelled the row. Opened on 3 September 2026, the Artificial Analysis image-to-video board lists the entry as Minimax H3 Max (post-trained by fal) at an Elo of 1,201 over 5,469 samples. fal's own page still says it is listed there under its internal name, MiniMax H3 Turbo (768p), with the note that its information was updated on 25 August 2026. Both statements were true when written; only the board has moved.

What is a MiniMax H3 Turbo LoRA?

It is a community label rather than an official release from MiniMax or fal. It refers to running the open-weights MiniMax H3 base model for roughly four to eight denoising steps instead of the usual twenty-five or more, using a step-distilled LoRA adapter or an acceleration framework, normally through ComfyUI on your own GPU. It speeds up the base model. It is not the hosted MiniMax H3 Max build and it does not produce the same weights.

Can I download MiniMax H3 Max?

Not as of 4 September 2026. The base model is open weights and published at MiniMaxAI/MiniMax-H3 on Hugging Face. Neither post-trained build fal produced from it — MiniMax H3 Max or H3 Max Turbo — has been released, and neither MiniMax nor fal has stated that either will be. Anyone offering MiniMax H3 Max weights for download is offering something else.

Is MiniMax H3 Max really the number one image-to-video model?

It is first on the Artificial Analysis image-to-video board with audio, at an Elo of 1,201 with a 95% confidence interval of minus 9 to plus 9 over 5,469 samples, read on 3 September 2026. Read the next two rows before quoting that. Dreamina Seedance 2.0 720p is at 1,190 and MiniMax H3 itself at 1,186, and at those intervals the top three overlap. First place is accurate; a decisive quality lead is not what the votes show.

What resolution and length does it produce?

480P or 768P, with 768P the default and the resolution it is tuned around — at 16:9 that is 1344 by 768 at 24 fps. Durations run from 5 to 15 whole seconds. Four seconds is rejected, which catches people moving over from the base model, and there is no 2K on this model at all. Audio is generated in the same pass as the picture rather than dubbed on afterwards.

Is MiniMax M3 related to MiniMax H3?

No. MiniMax M3 is MiniMax's language and agent model — a roughly 428-billion-parameter mixture-of-experts model with a one-million-token context window, released on 1 June 2026. It reads video but does not generate it. MiniMax H3 and MiniMax H3 Max are the video generation line. The two share a vendor and nothing else.

How fast is it, really?

fal reports a five-second 768P clip coming back in under three seconds on its own infrastructure. That is fal's figure on fal's hardware, and we quote it as theirs. We have not yet published our own measured times for this route, and until we have, treat any single number — including that one — as the vendor's claim rather than a specification.

Where can I run it without an API key?

Here. Text to video and image to video run the Max family in the browser with nothing to install, and the engine picker names which build a clip went to. What a clip costs is set out on pricing, and what the free tier covers is a page of its own because the honest answer needed more than a sentence.

I typed "fasth3" or "minmax h3". Same thing?

Almost certainly yes, and you want one of the two hosted builds. None of those are official names — there is no product called fast H3, H3max or MiniMax 3. The official strings are MiniMax H3 for the open-weights base model, and MiniMax H3 Max and H3 Max Turbo for the two post-trains fal hosts, plus the API identifiers listed in the table above.

Run the Model, Not the Name

Whichever of the three you came for, the two you can actually run are behind these pages, and the engine picker says which is which: 5 to 15 seconds, up to 768P, sound generated in the same pass as the picture.

Generate free · queue

Written and maintained by the MiniMax H3 Max AI Video Generator editorial teamPublished Last updated