Skip to contentOceanToken

7 min read

Hailuo Prompt Length: Limits and Keywords

Hailuo prompt length limits by model (2,000 characters on 2.3, 7,000 on H3), MiniMax's camera commands, words worth testing, and words to cut from prompts.

By OceanToken Team

Hailuo Prompt Length: Limits and Keywords

Editor's note: People copy 300-word prompts from forums because the clips look good, then wonder why theirs don't. We think the length was rarely the reason. A short prompt built from words that actually change the picture, tested one word at a time, beats a hoard of long ones. Use the extra room MiniMax gives you for choreography and dialogue, not for adjectives.

Short answer: Hailuo prompt length is capped at 2,000 characters on Hailuo 2.3 and 02, and 7,000 on H3 and H3 Max, per MiniMax's docs. Most shots need far less: a sentence or two covering subject, action, camera and light. Use bracketed camera commands such as [Push in] for precise moves, and save the extra length for multi-step action or dialogue.

Hailuo prompt length limits by model

These are the hard caps in MiniMax's own documentation, checked on 2026-11-09. The H3 figure comes from MiniMax's video generation guide , which gives one limit for both H3 models.

Model Prompt limit Where MiniMax states it
MiniMax-Hailuo-2.3 2,000 characters Text-to-video reference 
MiniMax-Hailuo-02 2,000 characters Text-to-video and first-and-last-frame  references
MiniMax-H3 7,000 characters Video generation guide
MiniMax-H3-Max 7,000 characters Video generation guide

The limits are in characters, not words or tokens. Validate length in your own code before you submit, and show the user a warning, rather than discovering the cap from an error after a job has queued. If your app lets users switch between 2.3 and H3, store the limit per model id, because a prompt that fits H3 can be more than three times the 2.3 cap.

Why shorter prompts usually win

Our rule of thumb, not a tested result: a long prompt is not wrong, but every extra subject or scene change is another thing the clip has to animate in a few seconds. A prompt that asks for a chef entering, tasting a sauce and plating a dish in one 6-second shot gives each action about two seconds, and appearance adjectives don't move at all. "A beautiful, stunning, elegant woman in a gorgeous flowing dress" spends a dozen words on a still image and none on motion.

Diagram: How long should this Hailuo prompt be?

There is a second reason. MiniMax's prompt_optimizer defaults to true, which means your prompt is rewritten before generation; the reference says to set it to false "for more precise control." With the optimizer on, a long, carefully worded prompt may not reach the model in the form you wrote. A short, clear one survives rewriting better. If you call Hailuo through OceanToken, note that the video guide  doesn't list prompt_optimizer, so write prompts that work whether or not they get rewritten.

Hailuo camera commands and prompt keywords

For camera movement, MiniMax documents something better than adjectives. On Hailuo 2.3, 02 and T2V-01-Director, you can write bracketed commands, and the reference says explicit commands "yield more accurate results" than free-form descriptions.

Move Commands
Truck [Truck left], [Truck right]
Pan [Pan left], [Pan right]
Push [Push in], [Pull out]
Pedestal [Pedestal up], [Pedestal down]
Tilt [Tilt up], [Tilt down]
Zoom [Zoom in], [Zoom out]
Shake, follow, static [Shake], [Tracking shot], [Static shot]

Up to three commands inside one bracket run at the same time, for example [Pan left,Pedestal up]; MiniMax recommends no more than three. Commands written in sequence run in order. For H3, the guide names [pan], [zoom] and [static] and says to place camera instructions directly after the description they apply to.

Speed and light are not documented as keywords, so treat these as words worth testing rather than a vendor list. For pace, "slowly", "gently" and "gradually" for calm shots, and a strong verb such as "sprints" or "accelerates" for intensity. Pick one; "fast" and "slow motion" in one prompt leave the model to choose. For light, short concrete terms like "overcast", "golden hour", "soft window light", "neon" and "backlit" change the frame more than quality words do.

Words to cut

These are working rules from writing prompts, not tested results, so treat each as a candidate for the swap test below. Quality boosters such as "8k", "masterpiece" and "ultra detailed" describe a wish rather than a scene. Stacked synonyms ("beautiful stunning gorgeous") repeat one idea three times. Instructions to an imaginary editor ("make it look amazing") give the model nothing concrete to draw. Negatives such as "no blur" name the unwanted thing; we'd describe what you do want instead. And a visible change ("her face goes from worried to relieved") is worth testing against a bare label like "sad".

A reusable Hailuo prompt skeleton

If you want a starting point rather than a blank box, use this order: subject and action, camera, environment, light. One sentence each is enough.

A courier in a yellow jacket locks his bike to a railing and checks his phone. Medium shot [Push in]. Rainy evening street, wet reflections. Soft neon light from a shop window.

That's 32 words, far below either cap. Everything beyond it should be a deliberate choice: add a second action only if the sequence needs it, add dialogue only on a model with audio, add style terms only if the look matters more than the motion.

When you start from an image, the prompt gets shorter still. The picture already carries subject, background and light, so describe only the change: what moves, how fast, and what the camera does. Re-describing the image invites the model to "correct" it. If the face and clothes must stay stable, say so once, in plain words.

When long prompts are right

Multi-step choreography and dialogue genuinely need words: sequence verbs, timing cues, quoted lines. That is what H3's 7,000 characters are for. Even then, splitting one long action into separate shots is often easier to control than one long prompt, because each clip has one job and you can re-run the one that fails.

Diagram: One long prompt split into three shots

If you'd like MiniMax to write the long version for you, the guide documents an H3-Context-IR task that returns an enhanced, structured prompt in content.prompt without generating a video. It's useful as a draft, but read it before you submit it, since it can add detail you didn't ask for.

Task Starting length Why
Single action, one camera move One or two sentences One subject, one motion
Image-to-video One sentence about the change The image carries the detail
Dialogue shot on H3 A few sentences with quoted lines Lines and timing need words
Multi-step choreography Split into several prompts Easier to control and re-run

These ranges are our proposed starting points, not vendor figures or measured results.

How to test this on your workload

We haven't published results for this article. To find the right length and words for your own shots, change one thing at a time and run each variant more than once, because a single run is noise.

Diagram: One-word swap test

Diagram: Prompt length ladder

  1. Length ladder: write one idea at four lengths (short, plus camera and light, plus environment, plus style adjectives) and run each several times.
  2. Word swap: keep the prompt fixed and change one speed or light word at a time.
  3. Log the word or length changed, the model id and settings, the number of runs, what changed on screen, and your confidence.

Our take: Write one subject, one action and one camera move, use MiniMax's bracketed commands for the camera, and spend extra characters only on sequence and dialogue. The caps are generous; good prompts rarely come close to them.

Running Hailuo prompts on OceanToken

On OceanToken the base URL is https://api.oceantoken.ai/v1, and /hailuo-api lists minimax/hailuo-3, minimax/hailuo-3-max and minimax/hailuo-2.3. The Hailuo 3 model page has the inputs and output resolutions for H3. The API is OpenAI-compatible, so a prompt test loop is a few lines with the SDK's videos.create_and_poll, and switching the same prompt from 2.3 to H3 means changing the model id, then your length check. One key and one prepaid balance also cover Kling, Seedance and Veo, which helps when you want to see whether a prompt problem is really a model problem. A failed video job is not charged; its hold goes back to your balance. For production polling, see our Hailuo API Python guide. See the current rate on /hailuo-api.

FAQ

What is the Hailuo prompt character limit? 2,000 characters on Hailuo 2.3 and 02, and 7,000 on H3 and H3 Max, per MiniMax's documentation.

Do camera commands work on H3? MiniMax's guide names [pan], [zoom] and [static] for H3, placed right after the description they apply to. The full list of 15 commands is documented for 2.3, 02 and T2V-01-Director.

Should I use negative prompts? Describe what you want instead. Negatives put the unwanted thing into the prompt and aren't a documented feature.

Does MiniMax rewrite my prompt? By default, yes: prompt_optimizer is true. MiniMax says to set it to false for precise control where the parameter is available.

Sources

  1. MiniMax API: text to video  (checked 2026-11-09)
  2. MiniMax: video generation guide (H3, H3 Max)  (checked 2026-11-09)
  3. MiniMax API: first and last frame video generation  (checked 2026-11-09)
  4. OceanToken docs: video guide  (checked 2026-11-09)
  5. OceanToken Hailuo API page (checked 2026-11-09)

Get your OceanToken API key

  1. Create an OceanToken account and an API key.
  2. Copy the snippet below, paste your key, and run it.
  3. Check the current rate on /hailuo-api.
curl https://api.oceantoken.ai/v1/videos \
  -H "Authorization: Bearer $OCEANTOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "minimax/hailuo-2.3", "prompt": "A courier in a yellow jacket locks his bike to a railing and checks his phone. Medium shot [Push in]. Rainy evening street, wet reflections."}'

Get API Key 

  • Hailuo
  • AI video
  1. 7 min read

    A Unified API for Chat, Image and Video

    Building on a unified API for chat, image and video: sync vs async calls, billing units, retries, timeouts and storage, with three calls and a launch checklist.

    • AI video
    • LLM API
    • Tutorial
  2. 7 min read

    Kimi K3 API International Access Guide

    Use the Kimi K3 API internationally: sign up with email, pay by card, create one key and call moonshotai/kimi-k3 with an OpenAI SDK. Steps, code and checks.

    • Kimi
    • LLM API
    • Tutorial
  3. 6 min read

    Call Tencent Hy3 With the OpenAI SDK

    Call Tencent Hy3 with the OpenAI SDK outside China: base URL, model id, streaming, tool calls, reasoning_effort values, Hy3 vs Hy3 preview, and token billing.

    • Hunyuan
    • LLM API
    • Tutorial