MiniMax compatibility

This API implements the MiniMax Video Generation V2 surface. The request and response bodies match platform.minimax.io field for field, so an existing integration moves by changing the base URL and the key.

Every difference is listed below. There are no others, and none of them change the shape of a request — only which values are accepted.

Differences

Item	Hosted MiniMax	Here	Why
`resolution`	`768P`, `2K`	`768P` only; `2K` → `400`	H3-Regenerate-2K, which produces 2K, is not part of the open-source release
`duration`	4–15	5–15; `4` → `400`	4 s is 96 frames, below the model's 120-frame floor
`model`	`MiniMax-H3`	plus `-Fast`, `-Economy`	Carries the service tier
`/v2/h3_context_ir`	present	not implemented	H3-Context-IR is a hosted multi-model workflow, not released
Cancel	no endpoint	`POST /v2/video_generation/{id}/cancel`	The status enum has `cancelled` but nothing produces it
`ref2va`	same service	separate worker pool; `503` if not deployed	It needs a different checkpoint
`seed`	not in the body	`X-H3-Seed` header	Keeps the body byte-identical to the official schema

On 2K and prompt expansion

Both are genuinely absent from the open-source release. 2K output comes from H3-Regenerate-2K and enriched prompts from H3-Context-IR; MiniMax has published neither.

We return a structured 400 naming the reason rather than silently downgrading a 2K request to 768p, or shipping a prompt expander whose output is not comparable to theirs. If you need either, MiniMax's hosted API still offers them, and their README describes running local H3-Base alongside those hosted endpoints.

The local-serving surface

MiniMax also publishes a contract for driving a local H3 deployment through SGLang:

POST /v1/videos          { task, prompt, conditions[], target, seed } → { id }
GET  /v1/videos/{id}                                                  → { status }
GET  /v1/videos/{id}/content                                          → mp4

We expose it too, authenticated and billed the same way, so a client already driving an SGLang box can point at us unchanged. The reference request bodies in MiniMax-H3/scripts/readme/*.sh are used verbatim as our test fixtures.

Internally this is the same surface our TPU workers speak, which is what makes an SGLang worker and ours interchangeable.

Using the official SDK

Point it at our base URL:

from minimax import MiniMax  # the official client

client = MiniMax(
    api_key=os.environ["H3_API_KEY"],
    base_url="https://api.h3.studio",
)

Everything in the video-generation namespace works. Text, speech and music endpoints do not exist here — this deployment serves H3 only.