Wan 3.0 allows 20,000 prompt characters, MiniMax H3 7,000

Alibaba's Wan 3.0 reference allows a prompt of up to 20,000 characters, while MiniMax's H3 reference caps each text item at 7,000. If you want one prompt to work on both wan-3.0 and minimax-h3 through Sume, write to the smaller limit, 7,000 characters, and spend the extra Wan room only when you pin Wan.

The numbers come from Alibaba's Wan3.0 Video Generation API Reference and MiniMax's Create Video Generation Task, both read on 2026-10-02. The Sume pages I read (Video generation and Video Router) do not state a prompt length cap, so I make no claim about one.

What are the two vendor limits exactly?

Alibaba lists a maximum prompt length of 20,000 characters for Wan 3.0 and turns prompt rewriting on by default. MiniMax's page says the text prompt is limited to 7,000 characters per item, and the request needs at least one non-empty text item in content.

Note the unit and scope: Alibaba's figure is the prompt, MiniMax's is per text item, with a 64 MB cap on the whole request body that matters when you inline media.

Does Sume tell me where it would stop me?

Sume's video guide says prompt is a required, non-empty string, and that a bad request is refused with 400 invalid_request or a more specific code. For parameters outside the catalog, such as size, seed or provider.options, Sume returns 400 unsupported_parameter rather than dropping them silently. A very long prompt is a case the docs do not spell out, so test the boundary once with your own model choice before you rely on it.

This also means you cannot lean on provider-side prompt expansion: Sume's provider.options list is empty in v1, so the rewrite switch Alibaba exposes is not reachable through the Sume route.

How do I keep one prompt portable?

A guard in your code is cheaper than a failed submission.

• Budget 7,000 characters for the whole text item and keep a 10 percent margin.
• Put the subject, action and camera in the first 1,000 characters; both vendors read the whole prompt, but short prompts are easier to review.
• Keep dialogue and sound direction in plain sentences, since both models take audio from the prompt.
• Store the prompt in your own system and send the same string for each model id you test.

import os, requests

LIMIT = 7000  # smaller of Wan 20,000 and MiniMax H3 7,000
prompt = open("shot.txt").read().strip()
if not prompt or len(prompt) > LIMIT:
    raise SystemExit(f"prompt must be 1-{LIMIT} characters, got {len(prompt)}")

for model in ("wan-3.0", "minimax-h3"):
    r = requests.post("https://api.sume.com/v1/videos",
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
                 "Idempotency-Key": f"portable-{model}-001"},
        json={"model": model, "prompt": prompt, "duration": 5})
    print(model, r.status_code, r.json().get("id"))

What should I do when a prompt needs more than 7,000 characters?

Pin wan-3.0 for that job and keep the longer text there, or split the idea into shots and generate one clip per shot. Be honest with yourself about whether the length helps: the extra characters are a limit, not a quality promise. Both vendors publish prompt guides, and Sume has a Wan 3.0 prompt guide and a multi-shot timecode post for the two styles.