MiniMax H3 Max

API model ID

minimax/h3-max/text-to-video

Access

POST https://queue.fal.run/minimax/h3-max/text-to-video with Authorization: Key <FAL_KEY>. GET returned status_url/response_url; PUT cancel_url to request cancellation. No authenticated generation performed.

Regions

Hosted API; inference residency unverified

Context / input limit

Not collected

Maximum input tokens

Not collected

Maximum output

Not collected

Reasoning

Not collected

Tool calling

Not collected

Structured output

Not collected

A JSON response containing a video file is not constrained JSON Schema generation.

Streaming

Not collected

Queue status/log streaming is not incremental video output.

Batch processing

Not collected

Queued requests do not establish a Batch Inference API.

Fine-tuning

Not collected

Public base weights and a public realism LoRA do not establish fine-tuning support for Max or Max Turbo.

Knowledge cutoff

Not collected

Release date

2026-08-26Source for Release date (opens in a new tab)

Announcement of the named Max and Max Turbo variants, not a later hosted route launch date.

Parameters

Not collected

Architecture

Not collected

Languages

Not collected

Context caching

Not collected

No documented prompt-cache controls, rates, minimum input or lifetime for these endpoints.

Maximum prompt characters

50,000Source for Maximum prompt characters (opens in a new tab)

fal API prompt limit; MiniMax native Max has a separate 7,000-character limit.

Output modalities

Video, AudioSource for Output modalities (opens in a new tab)

Synchronized video audio does not establish a standalone speech or music API.

Output resolutions

480P, 768P, 1080PSource for Output resolutions (opens in a new tab)

480P/768P native; 1080P latent refinement of a 768P source. MiniMax native Max supports only 480P/768P.

Video duration

0.92–15 seconds; fractional durations acceptedSource for Video duration (opens in a new tab)

fal request duration defaults to 5 seconds; actual output may be about 0.7 seconds longer. Use at least two seconds with target_audio_url.

Base model

MiniMax H3 open-weight base; post-trained by fal ResearchSource for Base model (opens in a new tab)

Input modalities

Text, AudioSource for Input modalities (opens in a new tab)

Required request fields

Non-empty prompt and prompt_expansion_mode are marked required. Expansion defaults to balanced in property metadata, but callers should provide it explicitly.Source for Required request fields (opens in a new tab)

Prompt expansion

disabled, balanced or quality; balanced is the declared default. Approximate expansion times are about one second or up to thirty seconds, not an end-to-end SLA.Source for Prompt expansion (opens in a new tab)

Random seed

Optional integer; omitted means random. Bounds are not specified.Source for Random seed (opens in a new tab)

Safety checker

enable_safety_checker defaults to true.Source for Safety checker (opens in a new tab)

Output transport

sync_mode=false returns a CDN URL; true returns Base64. This changes file transport, not whether inference is synchronous.Source for Output transport (opens in a new tab)

Timing measurements

timings.inference measures DiT denoising only, excluding the rest of the request. expanded_prompt may be null if unchanged, disabled or handled by the MiniMax hosted API.Source for Timing measurements (opens in a new tab)

Asynchronous generation

SupportedSource for Asynchronous generation (opens in a new tab)

Submit via queue.fal.run and poll the returned status/result URLs.

Cancellation and task deletion

Queued cancellation removes a request; in-progress cancellation signals the runner and may not stop execution.Source for Cancellation and task deletion (opens in a new tab)

Concurrent requests

Account-wide across endpoints: new accounts start at two concurrent running requests; self-serve capacity scales to forty with paid-invoice history. Per-model limits and enterprise allocations can differ. Inspect the account dashboard for actual capacity.Source for Concurrent requests (opens in a new tab)

Input media transport

Public URLs without required cookies/auth, presigned URLs, Base64 data URIs or SDK uploads. File uploads are public by default unless access controls are configured.Source for Input media transport (opens in a new tab)

Output link expiry

CDN lifecycle can be set at account or request level, including X-Fal-Object-Lifecycle-Preference. No universal output URL lifetime is assumed.Source for Output link expiry (opens in a new tab)

Soundtrack replacement

Optional HTTP(S)/Base64 audio, at least two seconds and at most 15 MB; requested output must be at least two seconds. Replaces the soundtrack, trims from the end or pads silence without changing playback speed. Exceptionally high input sample rates may be resampled to 96 kHz.Source for Soundtrack replacement (opens in a new tab)

Aspect ratios

21:9, 16:9, 4:3, 1:1, 3:4, 9:16Source for Aspect ratios (opens in a new tab)

Defaults to 16:9.

Prices retain their published currencies and billing units below.

API model ID

minimax/h3-max/image-to-video

Access

POST https://queue.fal.run/minimax/h3-max/image-to-video with Authorization: Key <FAL_KEY>. GET returned status_url/response_url; PUT cancel_url to request cancellation. No authenticated generation performed.

Regions

Hosted API; inference residency unverified

Context / input limit

Not collected

Maximum input tokens

Not collected

Maximum output

Not collected

Reasoning

Not collected

Tool calling

Not collected

Structured output

Not collected

A JSON response containing a video file is not constrained JSON Schema generation.

Streaming

Not collected

Queue status/log streaming is not incremental video output.

Batch processing

Not collected

Queued requests do not establish a Batch Inference API.

Fine-tuning

Not collected

Public base weights and a public realism LoRA do not establish fine-tuning support for Max or Max Turbo.

Knowledge cutoff

Not collected

Release date

2026-08-26Source for Release date (opens in a new tab)

Announcement of the named Max and Max Turbo variants, not a later hosted route launch date.

Parameters

Not collected

Architecture

Not collected

Languages

Not collected

Context caching

Not collected

No documented prompt-cache controls, rates, minimum input or lifetime for these endpoints.

Maximum prompt characters

50,000Source for Maximum prompt characters (opens in a new tab)

fal API prompt limit; MiniMax native Max has a separate 7,000-character limit.

Output modalities

Video, AudioSource for Output modalities (opens in a new tab)

Synchronized video audio does not establish a standalone speech or music API.

Output resolutions

480P, 768P, 1080PSource for Output resolutions (opens in a new tab)

480P/768P native; 1080P latent refinement of a 768P source. MiniMax native Max supports only 480P/768P.

Video duration

0.92–15 seconds; fractional durations acceptedSource for Video duration (opens in a new tab)

fal request duration defaults to 5 seconds; actual output may be about 0.7 seconds longer. Use at least two seconds with target_audio_url.

Base model

MiniMax H3 open-weight base; post-trained by fal ResearchSource for Base model (opens in a new tab)

Input modalities

Text, Image, AudioSource for Input modalities (opens in a new tab)

Required request fields

Non-empty prompt and prompt_expansion_mode are marked required. Expansion defaults to balanced in property metadata, but callers should provide it explicitly.Source for Required request fields (opens in a new tab)

Prompt expansion

disabled, balanced or quality; balanced is the declared default. Approximate expansion times are about one second or up to thirty seconds, not an end-to-end SLA.Source for Prompt expansion (opens in a new tab)

Random seed

Optional integer; omitted means random. Bounds are not specified.Source for Random seed (opens in a new tab)

Safety checker

enable_safety_checker defaults to true.Source for Safety checker (opens in a new tab)

Output transport

sync_mode=false returns a CDN URL; true returns Base64. This changes file transport, not whether inference is synchronous.Source for Output transport (opens in a new tab)

Timing measurements

timings.inference measures DiT denoising only, excluding the rest of the request. expanded_prompt may be null if unchanged, disabled or handled by the MiniMax hosted API.Source for Timing measurements (opens in a new tab)

Asynchronous generation

SupportedSource for Asynchronous generation (opens in a new tab)

Submit via queue.fal.run and poll the returned status/result URLs.

Cancellation and task deletion

Queued cancellation removes a request; in-progress cancellation signals the runner and may not stop execution.Source for Cancellation and task deletion (opens in a new tab)

Concurrent requests

Account-wide across endpoints: new accounts start at two concurrent running requests; self-serve capacity scales to forty with paid-invoice history. Per-model limits and enterprise allocations can differ. Inspect the account dashboard for actual capacity.Source for Concurrent requests (opens in a new tab)

Input media transport

Public URLs without required cookies/auth, presigned URLs, Base64 data URIs or SDK uploads. File uploads are public by default unless access controls are configured.Source for Input media transport (opens in a new tab)

Output link expiry

CDN lifecycle can be set at account or request level, including X-Fal-Object-Lifecycle-Preference. No universal output URL lifetime is assumed.Source for Output link expiry (opens in a new tab)

Soundtrack replacement

Optional HTTP(S)/Base64 audio, at least two seconds and at most 15 MB; requested output must be at least two seconds. Replaces the soundtrack, trims from the end or pads silence without changing playback speed. Exceptionally high input sample rates may be resampled to 96 kHz.Source for Soundtrack replacement (opens in a new tab)

Frame controls

First frame, last frame, both or neither. Canvas follows first, else last. Omitting both falls back to text-to-video at 16:9. No aspect_ratio field is exposed on this route.Source for Frame controls (opens in a new tab)

Prices retain their published currencies and billing units below.

API model ID

minimax/h3-max/reference-to-video

Access

POST https://queue.fal.run/minimax/h3-max/reference-to-video with Authorization: Key <FAL_KEY>. GET returned status_url/response_url; PUT cancel_url to request cancellation. No authenticated generation performed.

Regions

Hosted API; inference residency unverified

Context / input limit

Not collected

Maximum input tokens

Not collected

Maximum output

Not collected

Reasoning

Not collected

Tool calling

Not collected

Structured output

Not collected

A JSON response containing a video file is not constrained JSON Schema generation.

Streaming

Not collected

Queue status/log streaming is not incremental video output.

Batch processing

Not collected

Queued requests do not establish a Batch Inference API.

Fine-tuning

Not collected

Public base weights and a public realism LoRA do not establish fine-tuning support for Max or Max Turbo.

Knowledge cutoff

Not collected

Release date

2026-08-26Source for Release date (opens in a new tab)

Announcement of the named Max and Max Turbo variants, not a later hosted route launch date.

Parameters

Not collected

Architecture

Not collected

Languages

Not collected

Context caching

Not collected

No documented prompt-cache controls, rates, minimum input or lifetime for these endpoints.

Maximum prompt characters

50,000Source for Maximum prompt characters (opens in a new tab)

fal API prompt limit; MiniMax native Max has a separate 7,000-character limit.

Output modalities

Video, AudioSource for Output modalities (opens in a new tab)

Synchronized video audio does not establish a standalone speech or music API.

Output resolutions

480P, 768P, 1080PSource for Output resolutions (opens in a new tab)

480P/768P native; 1080P latent refinement of a 768P source. MiniMax native Max supports only 480P/768P.

Video duration

0.92–15 seconds; fractional durations acceptedSource for Video duration (opens in a new tab)

fal request duration defaults to 5 seconds; actual output may be about 0.7 seconds longer. Use at least two seconds with target_audio_url.

Base model

MiniMax H3 open-weight base; post-trained by fal ResearchSource for Base model (opens in a new tab)

Input modalities

Text, Image, Video, AudioSource for Input modalities (opens in a new tab)

Required request fields

Non-empty prompt and prompt_expansion_mode are marked required. Expansion defaults to balanced in property metadata, but callers should provide it explicitly.Source for Required request fields (opens in a new tab)

Prompt expansion

disabled, balanced or quality; balanced is the declared default. Approximate expansion times are about one second or up to thirty seconds, not an end-to-end SLA.Source for Prompt expansion (opens in a new tab)

Random seed

Optional integer; omitted means random. Bounds are not specified.Source for Random seed (opens in a new tab)

Safety checker

enable_safety_checker defaults to true.Source for Safety checker (opens in a new tab)

Output transport

sync_mode=false returns a CDN URL; true returns Base64. This changes file transport, not whether inference is synchronous.Source for Output transport (opens in a new tab)

Timing measurements

timings.inference measures DiT denoising only, excluding the rest of the request. expanded_prompt may be null if unchanged, disabled or handled by the MiniMax hosted API.Source for Timing measurements (opens in a new tab)

Asynchronous generation

SupportedSource for Asynchronous generation (opens in a new tab)

Submit via queue.fal.run and poll the returned status/result URLs.

Cancellation and task deletion

Queued cancellation removes a request; in-progress cancellation signals the runner and may not stop execution.Source for Cancellation and task deletion (opens in a new tab)

Concurrent requests

Account-wide across endpoints: new accounts start at two concurrent running requests; self-serve capacity scales to forty with paid-invoice history. Per-model limits and enterprise allocations can differ. Inspect the account dashboard for actual capacity.Source for Concurrent requests (opens in a new tab)

Input media transport

Public URLs without required cookies/auth, presigned URLs, Base64 data URIs or SDK uploads. File uploads are public by default unless access controls are configured.Source for Input media transport (opens in a new tab)

Output link expiry

CDN lifecycle can be set at account or request level, including X-Fal-Object-Lifecycle-Preference. No universal output URL lifetime is assumed.Source for Output link expiry (opens in a new tab)

Reference images

Up to nine images; at most twelve reference files across images, video and audio.Source for Reference images (opens in a new tab)

Input video limits

Up to three reference videos, each 2–15 seconds, combined video duration at most fifteen seconds.Source for Input video limits (opens in a new tab)

Reference audio limits

Up to three audio references, each 2–15 seconds, combined audio duration at most fifteen seconds. Audio-only references are explicitly allowed.Source for Reference audio limits (opens in a new tab)

Reference ordering and combinations

Use separate reference_image_urls/video_urls/audio_urls, or the ordered reference_conditions list, never both. Cross-modality order in the ordered list is semantic and preserved.Source for Reference ordering and combinations (opens in a new tab)

Frame controls

Optional first and last frames can accompany references. The first frame determines the canvas. A middle image requires both first and last frames and native 480P/768P.Source for Frame controls (opens in a new tab)

Middle-frame timing

Strictly between the first and last frames of the requested duration, rounded to the nearest frame at 24 fps; requires middle_image_url. Not available with 1080P refinement.Source for Middle-frame timing (opens in a new tab)

Aspect ratios

adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16Source for Aspect ratios (opens in a new tab)

Defaults to adaptive.

Reference video billing

4,096 shared free reference tokens per request, then USD 0.02 per 1,000 tokens, prorated. Apply the allowance once across all media. Output generation is additional and uses requested duration.Source for Reference video billing (opens in a new tab)

Reference image tokens

Allowed reference aspect ratios 0.4–2.5. Square images: 1,024 tokens (both 1024² and 2048²); 4:3/3:4: 1,376; 16:9/9:16: 1,824; 5:2/2:5: 2,560. The first four square images are free only if no other references use the shared allowance.Source for Reference image tokens (opens in a new tab)

Reference video tokens

Depends on aspect ratio, frame count and selected output resolution; 768P and 1080P share accounting while 480P uses fewer tokens. Counted frames are capped by requested output duration. Charges rise in steps, not a universal per-second rate.Source for Reference video tokens (opens in a new tab)

Reference audio tokens

Approximately eighty reference tokens per second; exact usage can vary. Audio embedded in a reference video is not separately counted as an audio reference.Source for Reference audio tokens (opens in a new tab)

Example reference cost

Five seconds at 768P with five square images and no video/audio references: 5 × $0.08 + (5 × 1,024 − 4,096) × $0.02 / 1,000 = $0.42048.Source for Example reference cost (opens in a new tab)

Prices retain their published currencies and billing units below.

API model ID

MiniMax-H3-Max

Access

POST https://api.minimax.io/v2/video_generation with Bearer API key and Content-Type: application/json; query GET /v2/query/video_generation/{task_id}. Pay-as-you-go API required.

Regions

International API; inference residency unverified

Request body limit

64 MB total request bodySource for Request body limit (opens in a new tab)

Input media transport

Public URLs, mm_file://{file_id}, or media-specific Base64 data URIs. URLs/platform files are recommended for large inputs; Base64 adds about 33% overhead.Source for Input media transport (opens in a new tab)

Asynchronous generation

SupportedSource for Asynchronous generation (opens in a new tab)

Returns a task_id; retrieve the result with the shared query endpoint. No inference call performed.

Task ID lifetime

Only tasks from the last seven days can be queried: [T−7d, T), with T in UTC seconds.Source for Task ID lifetime (opens in a new tab)

Task callbacks

Callback URL must echo the challenge within three seconds. Verified callbacks receive queued, running, succeeded, failed or cancelled updates.Source for Task callbacks (opens in a new tab)

Cancellation and task deletion

Queued tasks can be cancelled without charge. Running tasks cannot be cancelled. Succeeded/failed task records can be deleted; cancelled records cannot use this action.Source for Cancellation and task deletion (opens in a new tab)

Deleting a task record does not establish erasure of all provider-held inputs or outputs.

Input image formats

JPG, JPEG, PNG, WEBP, HEIC, HEIFSource for Input image formats (opens in a new tab)

Image input limits

≤30 MB per image; each side 256–5,760 pixels; aspect ratio 0.4–2.5. At most one first frame, one last frame, or nine reference images.Source for Image input limits (opens in a new tab)

Reference images

One first frame, one last frame, or up to nine reference images. First/last-frame roles cannot be mixed with reference roles.Source for Reference images (opens in a new tab)

Input video limits

At most three reference clips; each 2–15 seconds, total ≤15 seconds; ≤50 MB per clip. MP4/MOV; H.264/AVC or H.265/HEVC with AAC/MP3 audio; 256–5,760 pixels per side, aspect ratio 0.4–2.5, input frame rate 23.976–60 fps.Source for Input video limits (opens in a new tab)

Reference audio limits

At most three WAV/MP3 reference clips; each 2–15 seconds, total ≤15 seconds; ≤15 MB per clip.Source for Reference audio limits (opens in a new tab)

Input modalities

Text, Image, Video, AudioSource for Input modalities (opens in a new tab)

Output modalities

Video, AudioSource for Output modalities (opens in a new tab)

Maximum prompt characters

7,000Source for Maximum prompt characters (opens in a new tab)

Per text item; non-empty text required.

Output resolutions

480P, 768PSource for Output resolutions (opens in a new tab)

768P default; Max explicitly does not support 2K. Do not inherit fal 1080P.

Video duration

Integer 5–15 secondsSource for Video duration (opens in a new tab)

Four-second requests are unsupported for Max.

Required request fields

model=MiniMax-H3-Max, content containing non-empty text, resolution and integer duration. No Turbo selector documented.Source for Required request fields (opens in a new tab)

Aspect ratios

adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16Source for Aspect ratios (opens in a new tab)

Text-only requires a concrete ratio; first/last frames use adaptive and ignore other ratios; reference generation defaults to adaptive.

Prompt expansion

extra.prompt_expansion_mode: disabled, balanced or quality, default balanced. Other extra properties are not allowed.Source for Prompt expansion (opens in a new tab)

Reference ordering and combinations

First/last-frame image roles cannot mix with reference roles on this host. fal reference-route mixing is not inherited.Source for Reference ordering and combinations (opens in a new tab)

Output link expiry

Download URLs expire; query again for a fresh URL while the task is queryable within seven days. Exact URL lifetime is not stated.Source for Output link expiry (opens in a new tab)

rateLimit

Not collected

The family table names MiniMax-H3, not Max; no numeric Max-specific quota is inferred.

Context caching

Not collected

No native Max prompt-caching contract established.

Streaming

Not collected

Task completion and callbacks are not incremental video streaming.

Batch processing

Not collected

Fine-tuning

Not collected

Prices retain their published currencies and billing units below.