MiniMax H3 Max vs H3: Speed, Controls, and API Routes

Choose MiniMax H3 Max when turnaround matters more than output above 768P and your request is text-to-video or first/last-frame image-to-video. Choose MiniMax H3 when you need 2K output, mixed image/video/audio references, or downloadable base weights. H3 Max is not a blanket upgrade over H3. It is a fal Research post-trained variant of MiniMax H3, paired with an inference stack optimized for speed.^[1]

There is one more condition: the API route matters. MiniMax's direct API and the current reAPI route expose a narrower H3 Max contract than fal's own hosted endpoints. A feature shown on fal cannot be copied into a MiniMax or reAPI payload and expected to work.

Quick answer

Choose H3 Max for fast 480P/768P text-to-video or first/last-frame jobs; choose H3 for 2K output or MiniMax's broader reference inputs.^[3]
H3 Max is a fal post-trained branch of H3, not a speed toggle on the base model.^[1]
Capability depends on the endpoint: fal exposes an H3 Max reference route, while MiniMax direct and reAPI do not.^[4]^[7]
Base H3 has public weights; the cited first-party pages do not provide a separate downloadable H3 Max checkpoint as of September 7, 2026.

MiniMax H3 Max is a branch of H3, not H3 with a higher setting

The names invite the wrong mental model. “Max” sounds like the same product with every control turned up. The release history says something different.

MiniMax introduced H3 as a general-purpose multimodal model that understands text, images, video, and audio and generates video with native stereo sound. MiniMax then published H3 base checkpoints on August 3.^[2]^[6] fal Research started from those weights, added post-training data aimed at prompt adherence and visual quality, and co-optimized the resulting model with its inference system. The product of that work is H3 Max.^[1]

That lineage explains the trade. H3 Max targets a different point on the speed-and-control curve; it does not inherit every delivery tier and input mode on every service. The base remains the broader option in MiniMax's direct contract.

Use these names precisely in technical work:

Name	What it identifies	Do not infer
MiniMax H3	MiniMax's base H3 model and public base checkpoints	That every hosted H3 route has identical fields
H3 Max by fal	fal Research's post-trained H3 variant and optimized serving work	That “Max” adds 2K or every H3 control
`MiniMax-H3-Max`	Model value in MiniMax's direct V2 API	That fal's endpoint-specific fields are accepted
`minimax-h3-max`	Public model ID on reAPI	That it accepts base H3 reference arrays

MiniMax H3 Max vs MiniMax H3 specifications by route

The safest comparison begins with one surface, not one brand name. The table below separates the contracts that are live on September 7, 2026.

Capability	MiniMax direct: H3	MiniMax direct: H3 Max	reAPI: H3 Max	fal: H3 Max
Text-to-video	Yes	Yes	Yes	Yes
First frame	Yes	Yes	Yes	Yes
Last frame	Yes	Yes	Yes	Yes
First + last frame	Yes	Yes	Yes	Yes
Mixed reference-to-video	Yes	No	No	Separate endpoint
Output resolution	768P, 2K	480P, 768P	480P, 768P	480P, 768P
Duration	4–15 seconds	5–15 seconds	5–15 seconds	5–15 seconds
Public model weights	Base checkpoints published	No separate checkpoint verified	No separate checkpoint verified	No separate checkpoint verified

MiniMax's own V2 documentation is explicit about the central split: direct H3 Max does not support reference images, reference video, or reference audio, and it does not support 2K. Its image-to-video mode can take a first frame, a last frame, or both.^[3] The local reAPI request schema mirrors those limits and rejects extra fields rather than silently ignoring them.

fal's H3 Max reference route is real, but it belongs in the last column. Its current input form exposes reference image, video, and audio slots.^[4] Because endpoint coverage can change after launch, pin the endpoint and review date whenever this capability affects a product decision.

Resolution is the fastest way to decide

If the deliverable must come directly from the generator above 768P, MiniMax H3 Max is eliminated by its current contract. MiniMax direct H3 and the MiniMax H3 route on reAPI both expose a 2K tier. A later upscale is possible in a separate workflow, but it is not the same as asking H3 Max for a native, documented 2K option.

If the deliverable is 480P or 768P, resolution no longer settles the question. Turnaround, reference inputs, route availability, and acceptance rate take over.

Reference material is the second decision gate

A first frame fixes where a shot starts. A last frame fixes where it lands. Neither is equivalent to a reference pack.

Use the base H3 reference mode when the shot must combine jobs such as:

a portrait for subject identity;
a short video for movement or camera language;
an audio clip for voice or sound direction;
additional images for a prop, costume, or setting.

On MiniMax direct and reAPI, sending those base-model reference fields to H3 Max is an invalid request. If fal's reference-to-video route is the route you intend to buy and maintain, evaluate its own schema instead. Do not design a portable payload around the most permissive provider.

What the H3 Max speed claim does—and does not—establish

fal reported that H3 Max generated a five-second video in under three seconds of wall time on its optimized stack, with roughly 35 times the throughput of the official H3 endpoint in fal's comparison.^[1] That is meaningful evidence about fal's release and infrastructure. It is not a universal service-level guarantee.

MiniMax Design gives different user-facing figures: approximately 15 seconds for a five-second clip and 40 seconds for a 15-second clip, with variation by settings and service conditions.^[5] Those figures do not contradict the model being fast. They measure a different surface, potentially including work outside the optimized inference kernel.

A user's wait can include:

end-to-end latency =
  upload + admission + queue + prompt processing + inference
  + output processing + storage + polling delay + download

Only a matched test on the route you will ship can answer “how fast will it be for us?” The dedicated H3 Max real-time speed guide shows how to measure that full interval without relabeling a vendor benchmark as an independent result.

Quality claims need a matched acceptance test

fal says its internal human-preference evaluation ranked H3 Max ahead of the base across overall preference, prompt understanding, and aesthetics. The same announcement explains that fal selected checkpoints against its own head-to-head evaluations.^[1] That is a useful vendor result, not proof that H3 Max wins every prompt or preserves every H3 capability.

Run the model decision against the failures your project cannot tolerate. A compact test set should include more than scenic footage:

Timed action: three visible beats in a fixed order.
Camera plus subject motion: motion from both without geometry collapse.
Two speakers: distinct voices, correct speaker turns, and stable faces.
First/last interpolation: identical composition with a plausible path between endpoints.
Small fixed details: product geometry, garment marks, or a prop that should not change.
Your actual delivery ratio: include landscape and vertical if both ship.

Keep prompt text, duration, resolution, and source assets identical. Randomize which output reviewers see first, hide the model label, and write the pass conditions before generating. Then record two different numbers:

acceptance rate = accepted clips / completed clips

accepted clips per hour =
  completed clips per hour × acceptance rate

The second number can reverse a superficial speed result. A fast route that needs many reruns may deliver fewer usable clips than a slower one. Conversely, a modest quality difference may be irrelevant when H3 Max lets an editor test several concrete prompt revisions inside the same review window.

The reAPI request changes with the model choice

For H3 Max text-to-video on reAPI, send its exact public model ID, a concrete aspect ratio, and one of its two resolution values:

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer $REAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3-max",
    "prompt": "Five seconds, one continuous shot. A red paper boat moves through a shallow rain gutter as the camera tracks beside it. Natural rain and street ambience, no dialogue.",
    "aspect_ratio": "16:9",
    "duration": 5,
    "resolution": "480P"
  }'

The base-model payload becomes materially different when references are the reason for choosing H3:

{
  "model": "minimax-h3",
  "prompt": "Image 1 defines the character. Follow the pacing of Video 1 and use Audio 1 for the voice; keep the speaker identity stable.",
  "reference_image_urls": ["https://assets.example.com/character.jpg"],
  "reference_video_urls": ["https://assets.example.com/motion.mp4"],
  "reference_audio_urls": ["https://assets.example.com/voice.wav"],
  "duration": 6,
  "resolution": "768P"
}

Both jobs are asynchronous. Save the returned id, then poll the Tasks API until the state is completed or failed. Switching only the model string while leaving unsupported reference fields in place is not a valid migration.

Choose H3 Max or H3 by the constraint you cannot move

Your non-negotiable requirement	Better starting point	Reason
Fast 480P/768P prompt iteration	H3 Max	Speed is the purpose of the post-trained route
Direct 2K output	H3	H3 Max currently stops at 768P
Mixed references on MiniMax direct or reAPI	H3	Those H3 Max contracts do not expose R2V
H3 Max reference-to-video specifically on fal	H3 Max on fal	fal maintains a separate live R2V endpoint
Downloadable weights for local experiments	H3	Base checkpoints are published
Four-second output	H3	H3 Max starts at five seconds
First/last-frame shot at 768P	Test both	Both can express the shot; acceptance and route latency decide

This is also why an abstract winner is not useful. The better model is the one that satisfies the fixed constraint, then produces the most accepted clips inside the remaining time and budget.

FAQ

Is MiniMax H3 Max better than MiniMax H3?

Not for every job. H3 Max is the stronger starting point when fast 480P/768P turnaround is the priority. H3 remains the choice for direct 2K, four-second clips, broader reference inputs on MiniMax direct or reAPI, and downloadable base weights.

Who made MiniMax H3 Max?

fal Research developed H3 Max by post-training MiniMax H3 and pairing it with fal's optimized inference work. MiniMax made the base H3 model and published its base weights.^[1]^[6]

Does MiniMax H3 Max support reference-to-video?

It depends on the route. MiniMax's direct H3 Max API and reAPI's current H3 Max route do not. fal has a separate H3 Max reference-to-video endpoint. Name the provider and endpoint whenever you state this capability.

Can MiniMax H3 Max generate 2K video?

No, not through the current MiniMax direct or reAPI contracts. They expose 480P and 768P for H3 Max. Base H3 exposes 768P and 2K on those routes.^[3]

Is H3 Max open source because MiniMax H3 is open?

No. “Built from open weights” does not make the derivative weights public. Base H3 checkpoints are downloadable; no separate, verified H3 Max checkpoint appears in the first-party release materials cited here as of September 7, 2026. See the H3 Max open-source status guide for the evidence checklist.

Does H3 Max always return a five-second video in three seconds?

No. The sub-three-second figure is fal's reported result on its optimized stack, not a promise for every route, request, queue, or download. MiniMax Design publishes a different approximate user-facing time.

The shortest reliable decision rule

Start with MiniMax H3 Max for fast 480P or 768P text and frame-driven shots. Move to MiniMax H3 when the brief requires 2K, a four-second cut, mixed references on MiniMax direct or reAPI, or public base weights. If fal's H3 Max reference endpoint is part of the plan, record that route explicitly instead of describing reference-to-video as a universal H3 Max feature.

That one discipline—model plus provider plus endpoint—keeps most H3 Max versus H3 comparisons from becoming obsolete or technically wrong.

References

fal. Introducing H3 Max by fal. Published August 26, 2026. fal.ai
MiniMax. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities. Published July 31, 2026. minimax.io
MiniMax API. Create Video Generation Task V2. Retrieved September 7, 2026. platform.minimax.io
fal. H3 Max Reference to Video endpoint. Retrieved September 7, 2026. fal.ai
MiniMax Design. MiniMax H3 Max—Fast AI Video Generator. Retrieved September 7, 2026. design.minimax.io
MiniMax. Open General Intelligence: MiniMax H3 Is Now Open Source. Published August 3, 2026. minimax.io
reAPI. MiniMax H3 Max model page and request controls. Retrieved September 7, 2026. reapi.ai