MiniMax H3 Is Here: What Changed for Prompting, With 5 Real Week-One Prompts

MiniMax released Hailuo H3 on July 29, 2026, and the first wave of creator tests landed the same day. This post collects what is verifiable about the model so far, and — more usefully — the actual prompts people published in week one, with the structures that appear to be working.

Everything below is attributed to the creator or platform that published it. Where a number comes from a vendor or a distribution platform rather than independent testing, it is labelled as a claim.

What Actually Launched

The clearest technical summary came from fal, which <a href="https://x.com/fal/status/2083008450460541238" rel="nofollow" target="_blank">announced H3 availability on its platform</a>:

MiniMax H3 (Hailuo-03) is now available on fal. Open-weight multimodal video model with native stereo audio on every generation. Combines up to 9 images, 3 video clips, and 3 audio clips as references. Holds subjects, motion, and audio consistent. Weights will be released soon.

Those reference limits — nine images, three video clips, three audio clips — are the single most important number for prompting, and we will come back to them.

On specs and pricing, <a href="https://x.com/AndyMarlowg/status/2083218076699660442" rel="nofollow" target="_blank">AndyMarlowg</a> and <a href="https://x.com/Diana_Osire/status/2083131646702735618" rel="nofollow" target="_blank">Diana Osire</a> both reported the same three points via the Topview integration: native 2K resolution, 15-second clips, and audio and video generated together in one pass, at roughly 30% of Seedance 2.0's cost. Treat the pricing figure as a platform claim rather than an independent benchmark — it comes from the integration announcement, not from side-by-side testing.

The feature creators keep naming is Omni Reference. As <a href="https://x.com/VORTEX_Promos/status/2082505281590743519" rel="nofollow" target="_blank">VORTEX_Promos</a> described it, you can combine video, images, audio and text inside one prompt, and pin specific files with @ for tighter control.

The "Single Continuous Shot" Template Spread Fast

The most interesting prompting story of week one is not a feature — it is a template. Two different creators posted structurally identical prompts within a day of each other.

<a href="https://x.com/umesh_ai/status/2082499539735588916" rel="nofollow" target="_blank">Umesh (@umesh_ai)</a>, 297 likes:

Speeder chase across a cliff city (single continuous shot) From a monumental cliffside city carved into stone, the camera dives toward a tiny streak of light ripping along a narrow ledge-road. Lock-on: a speeder hugging the wall at insane speed. The camera slingshots ahead, whips back, then drops tight to the rear thrusters: heat haze, grit snapping off the ledge, warning lights flashing. A collapsing balcony rains debris; the rider snaps a last-inch swerve under a falling arch, then threads through hanging laundry lines and open windows in one fluid line.

<a href="https://x.com/Strength04_X/status/2082712692159344891" rel="nofollow" target="_blank">Strength04_X</a>, the next day:

Skyship flight across a floating kingdom (single continuous shot) From above a sea of endless clouds, the camera dives toward a majestic skyship soaring between gigantic floating islands. Lock-on: the airship accelerating into a violent thunderstorm suspended in the sky. The camera slingshots beneath the hull, whips across the towering sails, then drops tight behind the glowing engines: lightning tears across the clouds, rain lashes the deck, shattered rock fragments tumble through the air.

Strip the subject matter and the skeleton is identical:

1. Title line with the shot type in parentheses — (single continuous shot)
2. Establishing move — "From [wide vantage], the camera dives toward [subject]"
3. Lock-on: — an explicit instruction to attach the camera to the subject
4. Three-beat camera run — slingshots ahead / whips back / drops tight to [specific part]
5. Sensory detail bundled onto that third beat — heat haze, grit, lightning, rain
6. An obstacle sequence — something collapses, the subject threads through a gap
7. A reveal to end on — the camera bursts outward into a wide vista

This is worth stealing because it solves the two things video models fail at most often: it gives the camera an explicit path instead of a vibe, and it names an ending so the clip resolves instead of stopping mid-motion.

Native Audio Changes What Belongs in the Prompt

Because H3 generates stereo audio in the same pass as the video, sound is now part of the prompt surface rather than a post-production step. Several creators called this out as the practical difference from earlier Hailuo versions — <a href="https://x.com/ShamiWeb3/status/2083011335680585893" rel="nofollow" target="_blank">ShamiWeb3</a> described it as visuals, motion and audio arriving in a single creative pipeline.

The implication for prompt writing is simple: if you do not describe the sound, you are letting the model guess. Ambient bed, impact moments, and whether there is dialogue or only score are all worth one clause each.

Omni Reference: Nine Images Is a Lot

<a href="https://x.com/maxescu/status/2082834084343034181" rel="nofollow" target="_blank">maxescu</a> ran the reference ceiling to its limit and reported that it held:

Hailuo H3 is awesome for multi-character reference. I added 9 character sheets: Myself, Chris, Rourke, Linus, Lola, Kavan, Sebastien, Tim, and Jerrod. It did an awesome job keeping everything together.

Nine named characters in one generation is well past what most video models tolerate before faces start blending. If you are building anything episodic — a recurring cast, a brand mascot, a series with the same presenter — this is the capability to test first, because it is the one that historically forces people back into manual editing.

Negatives and Shot Lists Still Do the Heavy Lifting

For commercial work, the prompts that produced clean results were not the loosest ones. <a href="https://x.com/noorwithwifi/status/2082897758084907132" rel="nofollow" target="_blank">noorwithwifi</a> published this fashion spot prompt in full:

12s cinematic luxury fashion ad, 4K, 24fps, ultra-photorealistic. A high-fashion model in a flowing iridescent metallic gown walks along a minimalist desert runway at golden hour. Cinematic aerial, tracking, orbit, close-up, and hero shots with realistic cloth physics, flowing fabric, shimmering reflections, soft wind, golden sun flare, HDR lighting, Vogue-style editorial, seamless continuity, consistent face/outfit, no flicker, no warping, no extra limbs, fade to black.

Note the three-part structure: technical header (duration, frame rate, look), then the scene, then a run of constraints — consistent face/outfit, no flicker, no warping, no extra limbs, fade to black. The negatives are doing real work; so is naming the final transition.

A different approach from <a href="https://x.com/Maercihh/status/2082704405619679353" rel="nofollow" target="_blank">Maercihh</a>, who wrote a stop-motion assembly sequence as strict chronological steps — hands position twigs into an outline, then place bark for the face and jaw, then add layered strips for scale armor, then pack in moss. When the whole point of a clip is a process, sequencing the prompt as a process is what stops the model from cross-fading between the start and end states.

And <a href="https://x.com/lepadphone/status/2082586918525796858" rel="nofollow" target="_blank">lepadphone</a>, who had early access, made the case for the product-commercial use case specifically: one image, one simple prompt, a polished 2K spot — with the observation that what stood out was less the resolution than how well the model handled brand identity and the visual language of advertising.

What Carries Over

Nothing in week one suggests the fundamentals changed. Describe motion rather than restating what a reference image already shows; name your closing frame; use negatives to override the model's defaults instead of hoping a positive description wins. Our Hailuo H3 prompt guide covers that baseline structure in more depth, and it still applies — H3 mainly widens what you can feed in, not how you should write.

The honest caveat: this is week one. Every claim above is either a direct quote from the creator who ran the test or a platform announcement. Nobody has run controlled comparisons yet, and early-access impressions skew positive by construction.

Keep Up With H3 Prompts

We collect AI video prompts from X continuously and sort them by model, so the Hailuo and MiniMax examples accumulate as creators publish them. Browse them on trending prompts.

Working the other direction — you saw an H3 clip and want to know how it was built — is what our video to prompt tool is for. Paste a link or upload the clip and it returns a structured prompt covering scene, motion and camera work that you can adapt for H3 or any other model.