MiniMax H3 Max is API-only and its weights are not public, so run it here in the browser instead of on your own GPU.
Quick prompts
Video Preview
Ready for Generating
MiniMax H3 Max is a speed-tuned post-train of the MiniMax H3 video model, built by fal on the open H3 weights. It is a separate model, not a quality setting inside H3.
The name gets used three ways: max-quality settings on base H3, the MAX reference mode inside H3, and — this page — the standalone model fal serves over an API.
Its weights are not public. Base H3 opened in August 2026 and runs locally; this one does not. Using MiniMax H3 Max means calling a hosted API, which the workspace above does with no key and no install.
What the hosted route gives you that a local setup cannot.
The usual first step with an H3 workflow is checking whether your card has enough VRAM, and for many people the honest answer is no. Here there is nothing to check: a thin laptop and a 24GB desktop get the same file.
Running the H3 family locally means matching CUDA versions, adding attention backends, and replacing the default node graph before the speed is tolerable. This generator skips that layer: type the shot, choose a length and aspect ratio, generate.
Plenty of sites show a model name and quietly serve a smaller variant, a capped resolution, or a slower queue. We do not. Every render here goes to the MiniMax H3 Max text-to-video endpoint at 768P, for 5 to 15 seconds, at 30 credits per second — the same whether you generate once or a hundred times.
This workflow renders the whole scene from a text prompt; there is no opening reference image. Name the subject, give it one physical action, add one camera instruction, then one lighting detail. A single idea per generation makes two takes comparable.
Three names used interchangeably, and what each one is.
| Model | Output here | Clip length | Native audio | Credits per second | Available here |
|---|---|---|---|---|---|
| MiniMax H3 Max | 768P | 5-15 sec | No | 30 | Yes — the workspace above |
| MiniMax H3 (base) | 2K | 5-15 sec | No | 46 | Yes — on the MiniMax H3 page |
| MiniMax H3 Max Turbo | — | — | — | — | No — not offered here |
Turbo is a further speed-first variant fal publishes alongside this model; we do not route it. Choose base H3 when the deliverable needs 2K, and MiniMax H3 Max when you want more attempts for the same spend.
When a hosted 768P video model is the right call.
Threads asking whether H3 runs on 8GB of VRAM end with someone saying it technically loads, then takes hours per clip. Skipping that question is the point of a hosted MiniMax H3 Max generator: the render never touches your card.
Waiting forty-five minutes for a local render and getting an unusable take is the fastest way to stop experimenting. Moving the work off your machine changes what a bad take costs: you write the next prompt while the previous one runs.
CUDA upgrades, attention backends, a node graph rebuilt because the official one is too slow — an afternoon gone before a single frame exists. MiniMax H3 Max here has no install step at all.
A week of reviews that contradict each other is a common way to end up buying nothing. Running your own prompt through MiniMax H3 Max settles it for your material specifically.
Treat MiniMax H3 Max as the drafting layer: run many variations of a framing or camera move cheaply, then commit only the final version to a slower high-resolution model.
Vertical openers, background plates and filler cuts rarely need 2K, which is where 768P earns its keep. Pick 9:16 for a short-form hook or 21:9 for an establishing plate.
Three steps from idea to downloaded clip.
Step 01
Select MiniMax H3 Max in the workspace above and write the subject, action, camera move and mood. One idea per generation.
Step 02
Choose 5 to 15 whole seconds, then one of six aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Output is 768P.
Step 03
Start the render; the clip appears in the workspace to download. Sign in first for a file without a watermark.
The exact limits of the configuration this page calls.
| Specification | MiniMax H3 Max on this page |
|---|---|
| Input | Text prompt only |
| Output resolution | 768P |
| Clip length | 5-15 seconds, whole seconds |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Native audio | Not produced by this model |
| Cost | 30 credits per second (5s = 150, 15s = 450) |
| Watermark | On guest renders; none once you sign in |
| Commercial use | Supported |
These limits come from the configuration this page actually calls, so the MiniMax H3 Max duration ceiling really is fifteen seconds. Generation time is not listed because we have not finished measuring it, and an unverified number would be worse than none.
Weights, hardware, duration, cost and audio.
A post-trained version of the MiniMax H3 video model, produced by fal from the open H3 weights and tuned for stronger prompt adherence and higher throughput. It is its own model, not a quality setting inside H3.
Here it is concrete: 768P at 30 credits per second, against 2K at 46 for base H3. Use this one while exploring the shot; use base H3 when resolution matters.
Not currently. MiniMax opened the H3 base weights in August 2026, but the post-train has not been released, so it is reachable only through a hosted API — which this page gives you without a key.
People do get base H3 loaded on 8GB cards, usually with heavy offloading and times measured in hours per clip. Running it here removes the question: your card is not involved.
Five to fifteen seconds, in whole seconds — fifteen is the maximum MiniMax H3 Max duration here. Anything longer is several clips joined, and identity usually drifts at each join.
Not for this model itself — it is billed per second of output at 30 credits. The one free video Vogoo gives you each day runs a smaller model at 720p for about two seconds.
No. It is charged at 30 credits per second, so a 5 second clip costs 150 and a 15 second clip 450. Credits come from a plan or a one-time pack.
Turbo is a further speed-oriented variant fal publishes alongside it. We do not route Turbo, so this page cannot generate with it.
No. This route produces video with no audio track, so add sound in your editor. If you need native audio, use LTX 2.5 instead.
Six: 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Unlike base H3 there is no adaptive option, so a text-only request must name one explicitly.
Write one shot, choose a length and an aspect ratio, and render 768P video without touching a local install.
