MiniMax and Hailuo: What This AI Video Model Means for Your Ads

MiniMax Doubled at IPO. Its Model Makes Clips, Not Ads.

In January 2026, a Chinese AI lab most marketers had never heard of listed in Hong Kong, and its stock roughly doubled on the first day. The company is MiniMax. Its video model, Hailuo, is one of the cheaper good ones you can run right now. If you make ads, that sounds like your problem just got solved.

It didn't. A raw video model like Hailuo hands you footage. It does not hand you an ad. The gap between a clip that looks impressive and a clip that sells is the whole subject of this post. Here is what MiniMax and Hailuo actually are, what the recent news does and does not change, and where a raw video model fits the real job of making UGC ads.

Why MiniMax is suddenly everywhere

The name is showing up in feeds for two reasons at once: a loud financial event and a fast-moving model line. Both are real. Neither, on its own, changes what lands in your ad account this week.

The IPO that doubled on day one

MiniMax listed in Hong Kong in January 2026 and priced its offering at the top of its marketed range. Press coverage put the raise at least US$537.7 million, valuing the firm at around US$6.5 billion at that top price. Then the stock ran: it launched at HK$165 per share and reached peaks of HK$330, a rough doubling on debut. The company is backed by Alibaba and Tencent, and it builds a family of multimodal models: MiniMax M2, Hailuo 2.3, Speech 2.6, and Music 2.0. That is a well-funded lab with public-market money and a broad model roster, which is exactly why its releases now get amplified.

A model line moving fast

The video model is the part that matters most for ad makers, and it iterates quickly. Hailuo 02 introduced native 1080p output and 6 to 10 second clips with heavy attention to physics. Hailuo 2.3 built on that with, in MiniMax's words, "more complex character body movements with greater fluidity" and micro-expression changes that "are also more natural." MiniMax also folded its video tool into a broader Media Agent it pitches around "one-click video generation." The cadence is monthly-feeling, and it is meant to.

The price play

MiniMax competes on cost, out loud, and the receipts sit on its own pages:

• It calls Hailuo 02's pricing "the most open access and affordable pricing in the industry."
• It says Hailuo 2.3 "sets a new global record for video model cost-effectiveness," while holding the price steady.
• A faster variant, Hailuo 2.3 Fast, cuts costs "for batch creation by up to 50%."
• When it launched its M2 text-and-agent model, it made the API "available for free for a limited time."

Those promos are time-boxed, and that M2 window was tied to a launch, so treat any "free tokens" headline as temporary. The durable signal is the direction, not the coupon: MiniMax wants to be the cheap option, everywhere, and it is willing to lose money briefly to get there.

What Hailuo actually is, and is good at

Strip away the ticker and Hailuo is a video generation model. You give it a prompt or an image; it returns a short clip. Understanding its shape is the fastest way to see where it stops.

A raw text-to-video engine

Hailuo takes text or an image in and produces motion out. MiniMax has leaned hard into physical realism, branding Hailuo 02's motion "Extreme Physics Mastery," and the 2.3 update pushes body movement and facial performance further still. For a specific kind of shot, a gymnast mid-rotation, a product tumbling through water, a fabric catching wind, that realism is genuinely useful and hard to fake. This is real capability, not marketing air.

The specs that shape an ad

Two numbers set the practical envelope. Clips run 6 to 10 seconds, and output tops out at native 1080p. That is enough for a hook or a single beat. A finished UGC ad usually chains several beats and runs longer, which means the moment you want a real ad you are already assembling, not just generating. The model gives you the raw seconds; the sequencing, pacing, and message are still yours to build.

Where the realism helps, and where it doesn't

The trap is assuming better physics equals better ads. It depends entirely on what the ad is asking the footage to do. Hailuo's realism earns its keep when the shot itself is the message:

• A product surviving a stress test: dropped, splashed, or crushed on camera.
• A texture moment a phone struggles to catch: cream swirling, fabric moving, foam forming.
• A stylized world (anime, illustration, ink wash) that a real shoot could never stage.

It earns almost nothing when the ad depends on a person the viewer trusts:

• A creator looking into the lens and vouching for the product in their own words.
• A voice with a local accent that sounds like your buyer's neighbor, not a narrator.
• A claim delivered with the timing, breath, and cadence of real speech.

The first list is where a raw generation model shines. The second is where most UGC ads actually live, and it is exactly where generation alone leaves you stranded.

Cheap and fast, on purpose

The cost story is real, and it is the point. A model priced to win on volume lets you generate many takes instead of nursing one. That is the right raw material for a testing-heavy ad operation, where the economics reward volume over one perfect clip. Cheap generation is a genuine advantage. It is just an advantage in the ingredient, not in the dish.

Clips versus ads: the job a raw model skips

Here is the distinction the IPO headlines blur, and it is the one that decides whether Hailuo solves anything for you.

A model hands you footage

Hailuo, like every raw video model, outputs footage. Footage is an ingredient. An ad is a finished object with a job to do: stop the scroll in the first three seconds, make a claim, and move a specific person to act. A beautiful clip that does none of those things is a demo, not a creative you would put budget behind.

What a UGC ad needs on top

Turning a clip into a UGC ad means adding everything a text-to-video model leaves out:

• A script built around a hook and a specific angle.
• An actor whose accent sounds native to the audience you are selling to.
• A synthetic voice matched to lip movement, so it reads as a person and not a puppet.
• Burned-in captions for silent autoplay.
• A 9:16 format ready to drop straight into a campaign.

If you want the extreme version of this, ads that never show a face still need every one of those layers except the on-camera actor. None of it comes out of a generation model by default.

A worked example: one skincare ad

Say you sell a serum and want a testimonial ad. Hailuo can hand you a gorgeous 8-second clip of the bottle catching morning light, or a stylized hero shot no camera could get. What it will not hand you is the 25-second arc that actually converts: a person with your buyer's accent saying "I had breakouts for years," a beat that shows the product doing its job, a specific claim, and a call to action, captioned and cut to vertical. To ship that ad you still write the script, cast the actor, match the voice to the mouth, add the captions, and format it. Then you do it nine more times with different hooks, because the first version rarely wins. The generation was the five-minute part. The ad was everything that came after.

Volume is the real product

Then comes the part that actually moves paid social: not one clip, but many variations you can run side by side until one beats your benchmark. A raw model makes each clip cheaper. It does nothing, on its own, to turn those clips into a comparable set of ads you can launch, read, and scale. That translation from footage to a testable ad set is the work, and it is the same work whether the underlying engine is Hailuo or anything else.

The three ways teams misread a cheap model are worth naming, because each one wastes the very advantage cheap generation gives you:

• They confuse a demo for a creative. A clip that impresses in a group chat has not been asked to sell anything to a stranger.
• They generate one hero and stop. Paid social rewards the tenth variation, not the prettiest first one, so a cheap model with no testing loop is a Ferrari with no road.
• They skip the local accent. A native-sounding voice builds trust a neutral one cannot, and no raw model adds it for you.

How Hailuo compares to the models Novoads runs

The honest comparison is not Hailuo against one rival. It is a raw model against the workflow you wrap around any raw model.

Same category, different job

Novoads does not run Hailuo. It runs Seedance 2.0, Kling v3 Pro, Sora 2, Sora 2 Pro, and Google Veo 3.1 for video, plus image models like GPT Image 2 and Nano Banana Pro. Those are peers to Hailuo as raw engines: fast, improving on a monthly rhythm, and priced within a few dollars of each other. If you want to see how a couple of them stack up on their own terms, our Seedance versus Kling and Seedance versus Sora breakdowns go model to model.

The distinction that matters

A spec sheet will not tell you which raw model to marry, because the answer changes every few weeks. The useful comparison is the vertical one: what a raw model gives you versus what a finished ad requires.

Dimension	Raw video model	Ad workflow
Input	A prompt or an image	A product and a script
Output	A short clip	A finished 9:16 ad
Voice and lip-sync	Not included	Built in
Captions and format	Manual	Automatic
Testing loop	You assemble it	Variations by design
Best for	A single beat	Ads you can run and scale

What actually decides your pick

When raw models sit this close on quality and price, the model is the least interesting variable in the whole decision. The tiebreakers live in the workflow wrapped around it:

• Does the tool give you a native-accent actor, not just a talking face?
• Does it output captions and a 9:16 cut without a second app?
• Can you generate the tenth variation as easily as the first?
• Does your cost per test stay in single dollars, so testing stays guilt-free?

Answer those and it barely matters whether the engine underneath is Hailuo, Seedance, or Veo. Answer none of them and the cheapest model in the world still will not produce an ad you would put budget behind.

Engines are swappable; the workflow is not

The models trade places constantly. Seedance answers Hailuo, Veo answers Seedance, and next month something answers Veo. Our what Seedance 2.5 changes note is already one snapshot of a moving target. What survives every one of those handoffs is the workflow, the parts that never show up in a model's release notes:

• The brief and the specific angle you are testing.
• The script and its first-three-seconds hook.
• The actor and the accent your audience actually trusts.
• The vertical format and the burned-in captions.
• The number of variations you can afford to run.

Pin your operation to that list, and each new model becomes a free upgrade rather than a migration you dread.

What ad makers should take from the news

If you buy or build on AI creative tools, MiniMax's IPO is a postcard from your supply chain. Three practical readings, each a stance worth holding.

• Bet on workflows, not vendors. Your ad account does not care which lab made the model. Structure production so the engine is a swappable part and the brief, actor, and measurement stay fixed.
• Let the price war help you. MiniMax pushing "record" cost-efficiency pressures every rival to match it. Cheaper generation across the board means cheaper tests for you, no matter whose model you end up rendering on.
• Judge a tool by shipped ads, not IPO pops. A doubling stock and a viral demo are not a line in your ROAS report. The only question that matters this month is whether a tool ships converting ads this month.

How Novoads turns a model into an ad

Novoads is built on the exact assumption this story rewards: models are engines, and the workflow is the product. You write or auto-generate a script, pick an AI actor whose age, gender, and accent match your audience, and it produces a UGC-style vertical ad with voice, lip-sync, and captions, ready for TikTok, Reels, and Meta. The engine underneath is a detail the platform manages, not a decision you have to relitigate every time a lab ships.

Today it runs Seedance 2.0, Kling v3 Pro, Sora 2, Sora 2 Pro, and Google Veo 3.1 for video, and engines get added or retired as they earn their place. Hailuo was never one of them, so MiniMax's news does not touch the platform. Because each ad costs a few dollars rather than the $200 to $500 of a hired creator, roughly $2 to $11 depending on the model, the whole point is volume: spin up ten variations of a concept in a sitting, then let the data pick the one to scale. You can try that workflow end to end in Novoads. Start for $1: the trial gives you 3 days of access, then continues at $49/mo. Cancel anytime.

A clip is not an ad

MiniMax is a genuinely strong lab having a genuinely big moment, and Hailuo is a genuinely good video model. None of that makes it an ad tool. The doubling stock, the record-cheap pricing, the sharper physics, all of it lives one layer below the thing that actually converts. A raw model rents you better footage every quarter. What you own, and what turns whichever model is best this month into ads that ship, is the workflow around it.