Minimax H3 ships native 2K video with stereo audio baked into the same generation pass, holds multiple shots consistent inside one clip, and takes up to 12 separate reference files to guide a single output. Here’s everything it actually does.
What is Minimax H3
Minimax is the lab. Hailuo is the consumer app and brand name the model ships under, and coverage of this release uses “Minimax H3” and “Hailuo 3.0” interchangeably, so both names point at the same model. Worth clearing up before anything else, since a reader searching one name and finding coverage under the other shouldn’t assume they’re different products.
The headline spec, and the thing every source leads with, is native 2K output, a real jump from Hailuo 2.3’s resolution tier, paired with a longer native clip length and a considerably deeper reference system than the previous version offered.
Minimax H3 features
Generation modes: T2V, I2V, and Omni Reference. Three confirmed modes: text-to-video from a written prompt alone, image-to-video with support for both a first-frame workflow and a first-and-last-frame workflow, and Omni Reference mode, which functions as Minimax H3’s reference-to-video system, guiding a generation off multiple provided assets rather than a single starting image.
Resolution. Up to 2K, specifically a 1440-pixel short edge. A 768p mode has been announced as rolling out, presumably for faster or cheaper generation when full 2K isn’t needed.
Frame rate. 24 FPS, flatly, described consistently across sources as the cinema-standard frame rate rather than a variable or configurable setting.
Clip length. 5 to 15 seconds natively. A dedicated Extend Video tool pushes total output up to 30 seconds, separate from the native generation length, the same pattern several competing video models use, a base generation plus a distinct extension mechanism rather than one model natively producing a longer clip.
Aspect ratios. Six: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Omni Reference mode adds an auto-select option that picks a ratio based on the provided references rather than requiring one specified upfront.
Multishot generation. Confirmed, and worth treating as a real feature rather than a footnote: Minimax H3 generates multiple shots inside a single clip while holding lighting and appearance consistent across the cuts. That’s a meaningfully different claim than a single continuous take, it’s the model handling what would traditionally be an edit, a cut from one framing to another, inside one generation request.
Prompting ease. Genuinely mixed, and worth stating honestly rather than pretending it’s effortless in either direction. Professional shot language, phrases like “slow orbit to the right,” produces measurably more consistent results than plain scene description on its own. That means Minimax H3 rewards a bit of cinematography vocabulary. It’s not purely natural-language-friendly the way a model built to interpret casual description would be, closer to directing a shot than describing a scene.
AV sync. Native and strong, confirmed consistently across every source checked. Stereo dialogue, sound effects, and room tone generate in the same pass as the video itself, no separate audio production step required for short-form output.
Multilingual. Not explicitly confirmed anywhere in the official-adjacent materials checked for this piece. Voice cloning and voice transfer are both confirmed capabilities, real and specific. Multilingual dubbing or lip-sync specifically, the kind of claim Seedance 2.5 makes with a stated roughly-ten-language range, isn’t documented for Minimax H3 in any source reviewed here. Treat this as unconfirmed rather than assuming parity with competitors that do publish a language count, and test it directly against a specific non-English use case before relying on it.
AI video editing. Confirmed, instruction-based. Name the one element that needs to change, a character, an object, a background, lighting, an effect, dialogue, or voice, in natural language, and the model applies that specific change while holding the rest of the frame stable. This is the same underlying philosophy region-level editing takes in competing models: a targeted fix instead of a full regeneration.
Minimax H3 use cases
Film and advertising production, where the multishot feature and native audio combine to produce something closer to a finished sequence than a single generated clip. Ecommerce product showcases, leaning on the reference system to keep a specific product consistent across several angles or contexts. Gaming and UI/UX demonstration content, mockup footage for an interface or a game concept before either exists in a shippable state. Social content generally, and vertical drama specifically, a genuinely distinct, named short-form video format worth explaining for anyone unfamiliar: serialized, vertical-video storytelling built for mobile-first platforms, increasingly its own production category rather than just “short video.” Motion design and stylized animation round out the list, both leaning on the style-consistency strengths covered below.
Minimax H3 input methods
Up to 9 reference images, 3 reference video clips, and 3 reference audio clips, 12 files total, in a single generation. Reference video and audio clips each run 2 to 15 seconds.
What each type of reference can actually guide, specifically: character appearance, performance, camera movement, composition, voice identity, visual atmosphere, narrative rhythm, sound design, and editing style. That’s a wide spread, references aren’t limited to “what does this look like,” they extend into pacing and audio character as well.
One nuance worth a direct callout: these references guide the generation, they don’t function as forced keyframes. The model interprets what’s provided rather than strictly reproducing it frame-for-frame, which matters for expectations, a reference image sets direction and style, it doesn’t guarantee the output matches it pixel for pixel the way a literal keyframe constraint would.
A note on realism vs. stylized video
Minimax H3 has a genuinely confirmed strength in anime and 2D stylized animation, including holding character design, line quality, proportions, color, consistently across a full action sequence rather than just a single still frame. That consistency-across-motion claim is the harder problem in stylized generation, and it’s specifically called out as working, not just implied.
The one clear, repeatedly cited prompting rule worth building into any workflow: pick one style and commit to it. “Studio Ghibli anime style” works as a direct, specific instruction. “Photorealistic anime” doesn’t, mixing two incompatible style targets confuses the model rather than blending them into something coherent. That’s a small rule with an outsized practical impact, since it’s an easy mistake to make when a creative brief itself is asking for a hybrid look.
Balance that against the realism side of the model: native 2K resolution and the full reference system are equally built for live-action-style, photorealistic brand and product work, not stylized output exclusively. Minimax H3 isn’t a stylized-animation specialist tool that happens to also do realism, it’s built across both, and the style-commitment rule applies either direction, a realism-focused prompt benefits from the same discipline, naming one coherent visual target rather than stacking contradictory descriptors.
Minimax H3 prompt structure
The strongest source of real prompt examples is Minimax H3’s own gallery, which pairs 20 public results from 16 different creators across 7 use cases with each result’s original prompt, genuine worked examples rather than invented ones. Pulling the pattern from that gallery and the broader prompting guidance available:
A cinematic, live-action-style prompt leads with shot language before scene content, camera movement and framing named explicitly, “slow orbit to the right around a parked motorcycle, chrome catching the streetlight,” rather than opening with a flat description of the scene alone.
A stylized or anime-style prompt names the committed style upfront and holds to it through the rest of the description, “Studio Ghibli anime style, a girl running through a wheat field, wind moving through the grass, warm late-afternoon light,” style established first, then action and atmosphere layered on without introducing a second, conflicting style descriptor anywhere else in the prompt.
A product or UGC-style prompt leans on the reference system directly rather than describing the product from scratch in text, a product reference image plus a short prompt describing the desired motion and setting, “the reference product rotating slowly on a marble surface, soft studio lighting,” letting the image reference carry the visual accuracy that a text description alone would struggle to guarantee.
Across all three patterns, the same two rules from the features section hold: lead with camera and shot language over pure scene description, and commit to one visual style rather than blending descriptors that pull in different directions.
Minimax H3 vs. Hailuo 2.3: the upgrade highlights
Resolution jumps to native 2K from 2.3’s tier, the single most-cited improvement across every source. Duration stretches to 15 seconds natively, longer than 2.3’s cap. The reference system is new territory entirely, mixed image, video, and audio references combined in one generation wasn’t available on 2.3. And voice transfer, cloning a specific voice identity into a generation, is a genuinely new capability rather than a refinement of something 2.3 already did.
Worth stating plainly: this reads as a real capability jump across resolution, duration, and reference depth simultaneously, not an incremental quality polish on the same feature set.
Minimax H3 vs. Seedance 2.5 vs. Gemini Omni
| Spec | Minimax H3 | Seedance 2.5 | Gemini Omni |
|---|---|---|---|
| Max duration | 15s native, 30s extended | 30s native, 180s beta | 10s |
| Resolution | Up to 2K | Native 4K | Not confirmed at the same tier |
| Reference capacity | 12 files (9 img, 3 video, 3 audio) | 50 inputs (30 img, 10 video, 10 audio) | Conversational multi-turn, not a fixed file count |
| Audio | Native stereo, same pass | Native, same pass | Native, same pass |
| Editing model | Instruction-based, name the element | Region-level re-render | Conversational multi-turn refinement |
The honest positioning: Minimax H3’s real edge is 2K-at-its-price-point economics and multishot generation inside a single clip, not raw spec superiority over Seedance 2.5, which still leads on both maximum resolution and maximum reference capacity by a wide margin. Gemini Omni’s structural difference, conversational editing that preserves continuity across chat turns rather than a one-shot generation with references attached upfront, makes it a different kind of tool entirely rather than a directly comparable spec sheet.
Minimax H3 pricing, access, and availability
API access through fal.ai runs roughly $0.073 to $0.120 per second for 2K output without a video input reference, a range the source itself frames as 46 to 67% lower than competing 720p solutions priced around $0.223 per second. Worth flagging that comparison as a self-reported claim from the pricing page itself rather than an independently verified benchmark, worth a direct spot-check against current competitor pricing before repeating it as neutral fact.
The Hailuo consumer app runs a freemium tier structure: Free at $0 with 20 credits and limited model access, Basic at $9/month for 900 monthly credits, Standard at $24/month for 3,200 credits, Pro at $45/month for 6,200 shared credits, Pro Max at $170/month for 24,000 shared credits, and Enterprise at custom pricing for organizations with higher-volume or specific compliance needs.
Two distinct access paths worth separating clearly: the Hailuo app for a consumer-facing, point-and-generate experience, and API access via fal.ai and likely other aggregators for developers building Minimax H3 into their own pipeline or tool.
FAQs
Does Minimax H3 support multilingual dialogue?
Not explicitly confirmed in official materials. Voice cloning and voice transfer are real, documented capabilities; multilingual dubbing specifically isn’t published as a spec the way it is for some competitors. Test directly against your actual language need before assuming support.
What’s the real difference between Omni Reference and a forced keyframe workflow?
Omni Reference guides the generation’s direction and style; it doesn’t lock the output to match a reference pixel for pixel the way a true keyframe constraint would. Expect interpretation, not literal reproduction.
Is the 15-second native limit a hard cap?
For native generation, yes. The separate Extend Video tool pushes total output to 30 seconds, a distinct mechanism from the base generation rather than a longer native clip length.
Is Minimax H3 better for anime and stylized work, or live-action realism?
Both are genuinely supported strengths, not one at the expense of the other. The deciding factor is prompt discipline either direction: commit to one style and don’t mix descriptors.
How does Minimax H3 compare to Seedance 2.5 on raw specs?
Seedance 2.5 leads on both maximum native resolution (4K vs. H3’s 2K) and reference capacity (50 files vs. H3’s 12). H3’s advantages are elsewhere: per-second pricing at its resolution tier and multishot generation within a single clip.