MiniMax H3: every input, one take
What it does, what it costs, how it compares, the four ways to run it, and the footage to judge it on.
What sets it apart
Three things separate H3 from the rest of the field: where it sits on the public leaderboards, the work it does that other models cannot, and the four ways it can be deployed. They are worth taking in that order — where the model can run matters once the output has earned the evaluation.
H3 sits at the top of the video category alongside Seedance 2.5. Both models are measured in the open, so the boards are checkable rather than asserted — and on image-to-video H3 leads outright.
First and second on Artificial Analysis image-to-video with audio (H3 Max and H3), and first of 48 models on Arena's image-to-video board. Third on Arena's video edit board, inside the error bars. Full numbers in the next section.
The strongest argument is not where H3 scores a point higher, it is the work that cannot be done elsewhere at all: editing finished footage, and recombining a camera language, a character and a voice from separate sources.
Audio and picture in one pass, 43 Elo ahead of Seedance 2.0 on video editing · omni-reference and editing on existing footage · text survives motion, the category's usual failure point · open weights.
From self-hosted base weights to the fully hosted API. Each rung has a different commercial shape — compute only, then tokens, then per-second, then fully managed.
768P at $0.08 plus regeneration at $0.05 is $0.13 a second — identical to generating 2K directly. There is no price penalty for climbing the ladder.
Arena publishes error bars. H3 leads image-to-video by seven points against ±11, and trails on video edit by 22 against ±26. Both are ties; saying so is what makes the rest credible.
Per-second comparisons only mean something with the resolution attached. Unqualified price claims get corrected in the first reply.
The base result goes back through the model with the original context, so detail comes from the source material rather than from a super-resolution guess.
The number that governs a real budget is the cost of takes that ship, not the list price of every second rendered.
Where H3 places, board by board
H3 is first on Arena.ai's image-to-video board — the largest public video evaluation running, at 2.1 million votes — and it is the only open-weights model in that board's top ten. On Artificial Analysis, MiniMax H3 Max and MiniMax H3 hold the top two places on image-to-video with audio, nineteen points covering the whole top five. Name the board and quote the interval every time: these are preference scores, they carry error bars, and they move as votes accumulate. Figures below were read between late August and October 2026; each card gives its own date.
First across 48 models on the biggest public video board there is, on 57,112 votes. Arena gives H3 a rank spread of 1–3, so treat the top three as unseparated rather than claiming a clear lead — fifteen points cover them, against intervals of ±5 to ±11. H3 is the only open-weights model anywhere in the top ten; every other model on this board is proprietary.
Bars scaled from an Elo floor of 1,000. MiniMax models take the top two places, and nineteen points cover all five — with intervals of ±7 to ±10, only the ends of that list are actually separated. Quote the interval, not just the rank.
* H3 Max is post-trained by fal from H3's open weights and served on fal's own stack. It runs to 768P and does not offer 2K, reference-to-video or editing.
Wan 3.0 leads at 1,187, ahead of Dreamina Seedance 2.5 at 1,151 and Gemini Omni Flash 1.1 at 1,139 — all outside H3's interval. H3 is fourth on 7,023 samples, still well clear of Seedance 2.0 at 1,076. Say fourth. Anyone who checks will see it.
Wan 3.0 leads at 1,156. Utopai X — post-trained from H3 — is second at 1,149, Dreamina Seedance 2.5 third at 1,143, H3 fourth at 1,137 and H3 Max fifth at 1,130. Twenty-six points cover the top five against intervals of ±9 to ±10, so this is a cluster rather than a ranking — and three of the five are built on H3.
Wan 3.0 leads at 1,414 ±26 and Dreamina Seedance 2.5 is at 1,410 ±26. H3's interval overlaps both, so the top three are not separated. Figures as read on 26 August — check the board before quoting them.
How it compares to Seedance
Seedance 2.5 launched at the end of July 2026 and is now the model H3 is most often evaluated against. It is a strong model. The honest read is that the two are built around different centres of gravity: Seedance 2.5 is a generator; H3 is a generator that also edits. Switch below to see the older Seedance 2.0, which is still what most published comparisons use.
The gap widens rather than closes as quality goes up, because 720p is where Seedance 2.5 stops. Going higher means a separate upscaling step and a separate bill. H3 generates 2K natively, or regenerates into it inside the original context for $0.05 a second.
This is the comparison the "less than a third per second" line rests on. It holds at 1080P and above. At 720P it does not — H3 2K is about 13% lower there, not a third.
Seedance 2.0 is a strong model and its single-shot output is mature. The distinction is what happens next. In real production a shot is almost never delivered from one generation — characters recur, packaging and on-screen text change late, backgrounds, lighting, motion and pacing get revised repeatedly.
Whether a model can pick up existing material and keep working on it decides whether it enters the pipeline at all. That is where H3 leads.
Product reference, scene, voice-over, music and end frame in one pass; six variants across three languages in a batch.
One SKU across variants and markets: product reference without drift, multi-angle demos, localised voice-over.
Character cinematics, skill demos and patch trailers where character reference, multi-shot and audio must lock together.
Generation, editing, audio, multi-shot and continuation under one API — one integration, one billing class, less vendor risk.
Run it inside your own environment: nothing leaves the infrastructure and there is no per-second meter. Check the regional note first.
Across all five, the same argument does the work: the footage already in hand becomes the input, and picture and sound arrive together rather than in two production steps.
All Seedance figures are taken from the official BytePlus ModelArk pricing page, retrieved 11 August 2026, and are published pay-as-you-go list prices. Per-second figures are BytePlus's own, derived from its five-second 16:9 examples with text or image input and no reference video; BytePlus states these are estimates and that actual billing follows the token count returned by the API. Rates differ by region, aspect ratio and reseller, and Seedance 2.5's API availability has been staged — verify with the provider before relying on any of it. Reproduced for factual price comparison only and not verified by MiniMax. The products are not equivalent, which is why every row names the resolution it is measured against. MiniMax prices are list, exclusive of tax, and subject to change; the authoritative source is platform.minimax.io. Nothing here is an offer, a quotation or a contractual commitment.
Watch it work
Twenty-six cases in three groups: footage edited in place, references recombined, and work generated from stills or text alone. Each card shows what went in, what came out, and the instruction that produced it. Every clip is real H3 output, with sound — turn it up.
Multi-asset joint reference
→ one continuous cut, with score
Seven separate assets enter one context. Six stills supply the people, wardrobe and locations; the reference video supplies the shot rhythm, the transition style and the music. H3 reads them together and returns a single edited sequence. This is the capability no competing model offers at all — not a better score on a shared benchmark, a different shape of input.
"Images 1 through 6, following the shot rhythm, transition style and music of example video 1 strictly."
Nineteen words. The references carry the brief, so the prompt does not have to.
Edit footage you already have
Video in, video out. One instruction applied to a finished shot — the half of the workflow most teams have not switched on, and the capability Artificial Analysis ranks H3 first on.
Six discrete changes, one pass, everything else untouched.
A newspaper becomes a green hardback, the chair becomes a red sofa, sunglasses are removed, a burning car is restored, a photograph becomes a black notebook, and a tree is added on the left.
The least spectacular clip here and usually the most commercially useful — build the set once, let it read as any city.
Keep the interior plate exactly as shot. Replace only the scenery visible through the window.
The argument for seasonal campaigns that reuse one shoot.
Relight the scene from day to night. Framing, performance and set stay where they are.
The input came from another model — H3 does not care where footage came from.
Replace the drink can with the branded can, change the lettering on the storefront sign behind, replace the snacks to match, and rewrite the spoken line so the dialogue still makes sense.
One set ships empty for the product page and populated for the campaign.
Add a figure in matching uniform on the left, moving in step with the others.
Eleven words. The whole instruction is the whole prompt.
Replace the cat in the video with a dog.
Two reference images, two targets named by position in frame. Nothing else moves.
Replace the child at the very back of video 1 with the golden retriever in image 1, and change the khaki jacket worn by the child at the far left of video 1 to the denim jacket in image 2.
Two videos in: the green-screen plate and a second clip supplying the world. Relighting is part of the same instruction.
Remove the green-screen background from video 1 and replace it with a fairy-tale setting like video 2. The background elements must match the character's movement in video 1 exactly. Adjust the lighting on the character so it matches the new background completely.
An effects layer that follows performance and escalates on cue — the UGC treatment applied rather than generated.
Add orange-yellow hand-drawn doodles like those in image 1 around the two people in video 1. As the two draw closer the doodles multiply, growing from small sparks into a very bright glow; when they kiss, the doodles take on pink strokes.
Audio in, audio and picture out. Lip movement resolves to the new line.
Change the line spoken by the woman in video 1 to the line in audio 1, and adjust the performance slightly to match.
Everything on this page is someone else's material. The clips that settle it are yours — your footage, your revision note, returned as H3 output, usually within a couple of days.
andrian@minimax.ioRecombine separate references
A character from one source, motion from another, a voice from a third. The assets do not have to have met before.
Real footage with game logic imposed on part of it, selectively, with physics intact.
Keep the architecture, pedestrians and overall environment of video 1 in real-world style, unchanged. Convert only the trees and cars into a voxel building-block style, following image 1. The motion paths of the pixelated objects must remain normal, and the shadows and transmitted light of the real scene must be preserved.
One still supplies who; the reference video supplies how they move.
The character in image 1 follows the movement, expression and performance rhythm of video 1 strictly.
Choreography nobody on your side has to shoot twice.
The character follows the movement of video 1, performing street dance; the character follows images 1 and 2.
Motion transfer across a change of subject the source never contained.
Imitate the movement following video 1. Fixed-camera wide shot; the three men in suits in the source video are replaced by three photoreal capybaras, following the original motion paths strictly.
Take a timbre from a reference recording and have a character speak lines that were never recorded.
The character speaks a new line. The voice timbre follows audio 1.
Generate from stills or text
No source footage. Included because these are the most requested formats — and because text and interfaces surviving motion is the category's usual failure point.
A 360-degree presentation of a black ergonomic office chair. Close-up of the breathable mesh back with airflow visualised. Engineering animation of the lumbar support. Demonstration of multi-axis adjustment.
The character follows image 1; the UI style follows image 2. Menu renders START NEW GAME / CONTINUE / SETTINGS / EXIT GAME correctly, equipment panel slides in, loading bar runs 0 to 100%, HUD assembles as the world loads.
Keep the gallery frame, palette, 3D figurine quality and layout of the original image unchanged, and add a lively sound effect as the type appears.
First person, eye level, handheld game camera. The frame simulates a player advancing around a military base with an assault rifle, sweeping the reticle across the passage ahead.
Two magicians perform a swap trick. When the smoke clears the colours of their suits have exchanged — but the colour of their gloves is unchanged.
Claymation style. A sprinting fox reaches the cliff edge and launches into a slow-motion heroic leap above a lava canyon, the camera sweeping beneath its belly in mid-air.
Four consecutive keyframes carried as a single campaign cut, holding wardrobe, talent and grade across every transition.
Five stills assembled into a title sequence with period grade, typographic reveals and an original score cut to the edit.
A single still expanded into a trailer beat: scale, camera move, score and voice arriving together.
A robot arm descends, picks up a red block, moves it forward and sets it down. Physical cause and effect, no reference material.
Seamless white cyclorama, product-led fashion spot. Proof of the vertical format and the clean-studio look for social buyers.
Three ways in, one model
Switch the resolution to see what 2K regeneration buys. It is not a super-resolution pass: the 768P result goes back into H3 with the original context, so detail is recovered from the reference material rather than invented.
{{ modeBody }}
Four rungs, one ladder
Open weights drive adoption; two hosted modules carry the quality. H3-Context-IR resolves what a raw prompt leaves unstated across the full multimodal context. H3-Regenerate-2K takes the base output back to 2K while still holding that context.
768P, community quality, compute cost only. Nothing leaves your infrastructure and there is no per-second meter.
768P at MiniMax quality, billed on tokens. This is usually what people mean when they say the model is not following instructions.
2K at MiniMax quality, billed per second of output. Regeneration expects a source that meets the 768P output spec — the usual reason a first call fails.
All three modules packaged, newest checkpoint, someone accountable when a generation fails.
The community licence for the open weights does not currently cover the EU, UK, South Korea or the US. Self-hosting in those markets runs through a formal licence request, which we prepare and submit with you — start that conversation early, because it sits ahead of deployment. Context-IR is not published as open weights.
List price, and what actually drives the bill
H3 meters concurrency rather than RPM. Higher concurrency and commitment pricing are available against a usage commitment — ask us for terms for your volume.
Against Seedance 2.0's published list, "less than a third per second" holds at 1080P and above. At 720P it does not — H3 2K is about 13% lower there, not a third. State the rung and the comparison stands up.
$0.08 plus $0.05 regeneration is $0.13 — the same as going straight to 2K — so regenerating only the takes users keep costs nothing extra, and every discarded take was billed at $0.08. Context-IR at roughly $0.016 a shot filters more of them out before anything renders.
List price per second tells you little on its own. Establish what share of your current generations actually ship, then measure the same thing on H3 with your own content — that ratio is what moves a real budget.
H3 Fast returns a 480P clip in roughly 1.2–1.5× its own duration, against minutes on the full model — a different order of latency, for eight-tenths of a cent more per second. Faces hold up; fine texture is softer. Reach for it wherever someone is waiting on the render — interactive tools, iteration passes, high-volume variants — then finish the takes that survive at 768P or 2K.
The five questions that come up
Tap to open the answer. These are the questions we are asked most often, answered plainly.