MiniMax H3 Context IR

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

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 task envelope or structured prompt text does not establish caller-defined JSON Schema constrained generation.

Streaming

Not collected

The reviewed contract describes asynchronous task completion, not incremental output streaming.

Batch processing

Not collected

Asynchronous tasks, lists and callbacks are not a documented Batch Inference API.

Fine-tuning

Not collected

Public H3 Base training code/weights do not establish fine-tuning of this managed component.

Knowledge cutoff

Not collected

Release date

Not collected

Parameters

Not collected

Architecture

Multi-stage preprocessing and orchestration using multiple hosted models and servicesSource for Architecture (opens in a new tab)

Internal checkpoint identities and aggregate parameter count are undisclosed in reviewed sources.

Languages

Not collected

Context caching

Not collected

No prompt-cache API, eligibility, cache lifetime or billing rate is documented for this component. Internal tensor/activation caching in H3 Base is not endpoint prompt caching.

API aliases

MiniMax-H3Source for API aliases (opens in a new tab)

Shared API selector for the H3 family; task route/type identifies this component. Not a separate base checkpoint ID.

Input modalities

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

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.

rateLimit

Not collected

The quota table gives Video Generation V2 / MiniMax-H3 300 RPM and 30 inflight tasks, but does not expressly scope that pool to IR or regeneration. No inherited numeric component quota.

Published quota scope

Published family table: Video Generation V2 / MiniMax-H3, 300 RPM and 30 inflight tasks. Component-specific applicability remains unverified.Source for Published quota scope (opens in a new tab)

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)

Maximum prompt characters

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

Per text item, measured in characters. A non-empty text item is mandatory; this is not a 7,000-token context limit.

Output modalities

TextSource for Output modalities (opens in a new tab)

Result is content.prompt for task_type=h3_context_ir, not video.

Required request fields

model=MiniMax-H3, content with one non-empty text item, and integer target duration 4–15 seconds. A concrete aspect ratio is required for text-only input.Source for Required request fields (opens in a new tab)

Downstream video duration

Integer 4–15 secondsSource for Downstream video duration (opens in a new tab)

Describes the downstream video plan; this component outputs prompt text, not a clip.

Aspect ratios

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

Text-only: explicit concrete ratio required. First/last-frame: adaptive; other valid ratio values are ignored. References: optional, defaults to adaptive.

Output formats

Text promptSource for Output formats (opens in a new tab)

Usage accounting

Successful tasks report prompt_tokens, completion_tokens and total_tokens. The managed workflow may use more billed tokens than the visible user prompt.Source for Usage accounting (opens in a new tab)

Reference-role discrepancy

Top-level scenarios allow a last-frame-only image, but the nested last_frame role description says it must pair with first_frame. Preserve both statements; last-frame-only validation is not confirmed by inference.Source for Reference-role discrepancy (opens in a new tab)

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

Why it is included

MiniMax explicitly identifies H3-Context-IR as a hosted, multi-stage system using multiple models and services, excluded from its public release. Included as a closed-source managed component; no independent checkpoint, private base weights, or parameter count is claimed. Public H3 Base checkpoints remain excluded.

Lifecycle dates

Sources reviewed