MiniMax H3 is here
2K video with real sound from anywhere
MiniMax H3 — the model everyone searches as “Hailuo 3” — generates up to 15 seconds of native 2K video with stereo sound, multi-shot scenes and dialogue in 11 languages in a single pass. Open weights are region-locked and capped at 768p. Renderforest runs the full hosted model: real 2K, any country, nothing to install.
What makes H3 different
Released July 31, 2026: 4–15 seconds of native 2K at 24fps, with 32kHz stereo generated in the same pass. H3 predicts picture and sound together and returns several shots from one prompt. Here is what that changes in practice.
Sound generated with the picture, not bolted on
H3 produces native 32kHz stereo in the same pass as the video — dialogue, footsteps, room tone, score. Not a voiceover track laid over a silent clip: the audio and the frames come from one model, so lip movement, impacts and ambience actually line up.
Native 2K, not an upscale
A rebuilt tokenizer (MiniMax calls it H3-VAE) gives the model roughly four times the effective sequence length, so 2K is generated rather than enlarged after the fact. Text on signage, logos and UI elements survive at full size — the thing most video models still get wrong.
Multiple shots from one prompt
Ask for a wide, a close-up and a reaction and H3 returns them as one coherent sequence with consistent characters and lighting — no stitching three separate generations and hoping the face matches.
Feed it up to nine references
H3 accepts up to 9 reference images, 3 video clips and 3 audio clips in a single context — 12 files in total — so you can hand it a character sheet, a location, a motion reference and a voice at once. Reference-driven direction instead of prompt roulette.
Dialogue in 11 languages
Speech is generated in-model with lip-sync across Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish — so a scene written in Portuguese or Arabic does not need a separate dubbing pass.
Edit by instruction, in the same model
Inside reference generation, H3 will transfer motion from a clip you supply and take plain-language edits on footage it already made — “make it night”, “keep the character, change the room”. One model covers what used to take three tools.
Open weights, region-locked. Here it just works.
Access status
- MiniMax H3 · released July 31, 2026 · 33B parameters
- Weights published on Hugging Face — August 3, 2026
- Self-hosting excluded in: United States, European Union, United Kingdom, South Korea
- Commercial self-hosting above $20M annual revenue requires written authorisation
- Open weights max out at a 768px short edge — the 2K module was not released
- Hosted model: full 2K, available worldwide — including in Renderforest
MiniMax published H3's weights under a community licence that names the United States, the European Union, the United Kingdom and South Korea as excluded territories for running the model on your own hardware, and requires written authorisation above $20M in annual revenue.
There is a second catch that almost no coverage mentions.
MiniMax held back two pieces of the pipeline from the open release: the prompt-processing module (H3-Context-IR) and the 2K regeneration step (H3-Regenerate-2K). Without the second one, a local install generates at a 768-pixel short edge. So the headline everyone read — “open-weight 2K video model” — describes the hosted model, not the one you can download. Nobody, in any country, is getting 2K out of the open weights today.
Here, you get the real thing.
Renderforest runs the full hosted pipeline, so you type a prompt and get genuine 2K with audio — from any country, on any device, with nothing to install and no GPU to rent.
And if you did want to run it locally
The weights are on Hugging Face, ComfyUI shipped native support on August 3, 2026 alongside the release, and AMD announced day-0 support on Instinct MI355X and MI300X. Worth doing if you are outside the excluded territories, have the hardware, and can live with 768p. If not, this page is the shortcut.
From prompt to published in four steps
- Step 1
Describe the shot — and the sound
Write what happens and what you hear. H3 responds to audio direction the way other models respond to camera direction: “rain on a tin roof”, “she says it quietly, in Spanish”, “bass drops on the cut”.
Describe the shot — and the sound - Step 2
Add references if you have them
Drop in a character image, a location photo, a clip whose motion you want copied, or a voice sample. Up to nine images, three clips and three audio files in one go.
Add references if you have them - Step 3
Generate — then compare
Run the same prompt on H3, Veo 3.1, Kling and Seedance side by side and keep the winner. Different models win different shots; one plan covers all of them.
Generate — then compare - Step 4
Finish and publish everywhere
Stitch your shots into a full video in the editor, add your logo, brand colours, subtitles and licensed music — then one click resizes for YouTube, TikTok, Reels and Shorts. Watermark-free on paid plans.
Finish and publish everywhere
Which job is H3 right for?
Match the capability to the work. Every one of these runs on the same plan, next to Veo 3.1, Kling 3.0 and Seedance 2.0.
Every H3 trick, one prompt each
Every clip below is MiniMax H3, and the prompt behind it is printed in full. Copy any prompt and run it yourself.
“Luxury coastal ad: giant [Sheewally] descends by parachute with realistic wind, weight and gentle sway. Soft landing, parachute collapses naturally. Reveal [50%] and [BUY NOW]. Keep product unchanged, sharp, upright and fully visible.”
“Bee-eye POV races through a sunlit meadow toward a realistic hive. Golden honey naturally forms the single-line title ‘MINIMAX H3’ across the honeycomb, while realistic bees fly around it. Hyper-realistic macro footage, cinematic light, 3–4 sec hold.”
“A blonde woman stands ankle-deep in turquoise surf wearing a flowing gown made of living seawater. Its liquid train ripples and blends into the waves as foam surrounds her feet. Soft hazy light, photorealistic fashion editorial, 9:16.”
“Cheerful anthropomorphic tomato chef cooks margherita pizza in a warm rustic Italian kitchen, adding mozzarella, basil and sauce before taking a cheesy bite with delight. He then smiles at camera holding a ceramic mug. Pixar-style 3D, cinematic, warm bokeh.”
“Five diverse colleagues gather around a conference table, actively discussing charts, laughing, gesturing and sharing ideas. Natural smiles, quick glances and playful energy. Bright modern office, soft daylight, photorealistic corporate style, 16:9.”
“3D doll-like barista makes an iced latte at a tropical tiki coffee bar during golden hour. She grinds beans, pulls espresso, adds ice and pours the coffee, then smiles at camera. Turquoise ocean, palms, warm sunset, glossy Pixar-style 3D.”
Everything about MiniMax H3
What MiniMax H3 actually is
MiniMax H3 is the omni-modal generation model MiniMax released on July 31, 2026. Give it text, images, video or audio and it returns video up to 2K with stereo sound made in the same pass — 4 to 15 seconds, at 24 frames per second, in aspect ratios from 21:9 through 9:16. Three modes cover the work: text-to-video, image-to-video with first- and last-frame control, and reference generation from up to nine images plus video and audio clips, which is where motion transfer and instruction-based editing live. At launch it placed first for video editing on the Artificial Analysis leaderboard, the first open-weight model to top an AI video ranking.
MiniMax H3, Hailuo 3 or M3?
MiniMax calls it MiniMax H3 — in the release notes, the research blog, the model card, the API and the Hailuo app itself. There is no official “Hailuo 3.0”. The confusion is fair: the two previous consumer video models really were Hailuo 01 and Hailuo 02, so a Hailuo 03 was the reasonable guess, and several platforms list H3 under that name anyway. If you searched for “Hailuo 3”, or for “hailou ai”, you are in the right place — it is the same model. What it is not is MiniMax M3: the M-series, from M2 to the M3 released June 1, 2026, are text language models. The video line is the H-series. Search “MiniMax M3” and you get a language repo.
Is H3 free, and usable here?
Three answers hide behind one question. The weights, published on Hugging Face on August 3, 2026, are free to download but region-restricted — MiniMax’s community licence excludes the US, the EU, the UK and South Korea from self-hosting — and they cannot reach 2K, because the regeneration module was never released; local generation caps at a 768-pixel short edge. The hosted API works worldwide at full 2K, but bills per second: about $0.13 a second, roughly $1.95 for one 15-second clip. Platform access like Renderforest’s works everywhere, gives the same 2K with audio, and adds a free tier — no GPU to rent, no arithmetic to do.
What changed since Hailuo 2.3
More than a version bump. Hailuo 2.3 generated 768p or 1080p in fixed 6- or 10-second clips (1080p only at 6 seconds), from a single source image, with no sound and no multi-shot. H3 generates 2K at any whole second from 4 to 15, takes up to nine reference images plus video and audio, produces 32kHz stereo alongside the picture, and returns several shots from one prompt with the same characters in each. In practice the silent single take has become a short scene with dialogue. Prompts written for 2.x still work — add audio direction and shot descriptions, and nothing you learned on the older model goes to waste. The 2.x habits carry straight over.
H3 next to Veo, Kling, Seedance
No single model wins every shot, which is the honest thing most model pages leave out. H3’s edge is generated audio at 2K with multi-shot structure and lip-synced dialogue in 11 languages, at a fraction of the per-second cost of the frontier alternatives. Veo 3.1 goes further on resolution and length, with 4K and scene extension past a minute. Kling 3.0 leads on character consistency across longer sequences and cuts up to six shots per generation, and Seedance 2.0 takes the most reference files. The practical approach is to run one prompt across all four and keep what works — which costs nothing extra when they share a subscription and an editor.
Why a clip is not a video
H3 caps at 15 seconds, and every real project runs longer than that. So the question that decides whether an AI video tool is useful is what happens after generation: can you cut four H3 shots together, drop your logo on, add subtitles in the language your audience reads, score it with music you are licensed to use, and export it at the right dimensions for three platforms? Renderforest was a video platform for a decade before it was a model menu — the generator feeds an editor, a template library and a brand kit, so your projects outlive whichever model is state of the art this quarter, and whichever one replaces it.
Loved by Creators
If you're looking for the best AI video creation tool, just try Renderforest. It's the only option that produces videos 100% true to your prompt and the reference images you add.
Short learning curve and very intuitive. I created a product explainer video a couple of hours after signing up, with no video editing knowledge.
I love using Renderforest for creating intro and AI videos. Everything is fast and works exceptionally well, especially the video quality.
With the AI tools they added, you can create a professional commercial for your business or brand just by typing a script.
I love the AI feature of Renderforest — that is just awesome. And I can make eye-catching reels using the AI reels maker.
Renderforest offers a diverse range of options for editing our videos with AI or manually. Unlike other AI tools, it's cost-effective and efficient.
The new AI features are awesome. Using image to video, I created beautiful short animations with my kids' book characters.
I don't come from a video production background, but the platform turns my scripts into clear, professional videos without any hassle.
A leap in the quality of our campaigns. With this speed of production, we serve retail clients with media on par with large ad agencies.
The AI features are very impressive. These tools make the creation process even faster and smarter, especially for video and branding content.
I've been using Renderforest for my video creation needs and I'm beyond impressed. The variety of templates is fantastic. A solid 5 stars!
The user-friendly interface, wide template collection, extensive customization options and fast rendering speed make my creative work achievable.
Hundreds of videos later, I'm continually impressed with the quality of the results. The best money I've spent in 30 years of marketing.
I'm dedicated to audiovisual creation, and Renderforest helps me create professional-level work quickly and simply. Fast, simple and pro.
One flat plan. MiniMax H3 and every model beside it.
- MiniMax H3
- Veo 3.1
- Kling 3.0
- Seedance 2.0
- Pixverse V6
- Renderforest Fast
Direct H3 API access runs ~$0.13/second at 2K — about $1.95 per 15-second clip, more after reseller margins. Renderforest bundles H3 with Veo 3.1, Kling, Seedance and Pixverse in one subscription: editor, 1200+ templates, brand kit, commercial rights. No credit packs, no per-second arithmetic — and the next breakthrough model joins at no extra cost.
Frequently asked questions
What is MiniMax H3?
MiniMax H3 is an omni-modal video generation model released July 31, 2026. It generates 4 to 15 seconds of native 2K video at 24fps with 32kHz stereo audio produced in the same pass, supports multi-shot scenes and lip-synced dialogue in 11 languages, and accepts text, images, video and audio as input. You can use it free in Renderforest without downloading anything.
Is MiniMax H3 the same as Hailuo 3?
Yes — one model. MiniMax's own name for it is MiniMax H3, used in the release notes, the model card and the API. “Hailuo 3” and “Hailuo 3.0” are unofficial names that caught on because the two previous consumer models were Hailuo 01 and Hailuo 02, and several platforms list H3 that way. It is unrelated to MiniMax M3, which is a text language model.
Can I use MiniMax H3 in the US, UK or EU?
Yes, through a hosted platform. MiniMax's open-weights licence names the US, EU, UK and South Korea as excluded territories, but that restriction is on running the weights on your own hardware. Using H3 through Renderforest works from any country — and gives you the full 2K pipeline, which the downloadable weights cannot produce.
Is MiniMax H3 free?
The weights are free to download where the licence permits, but MiniMax withheld the 2K regeneration module, so a local install caps at a 768-pixel short edge. The official API is paid — about $0.13 per second at 2K, $0.08 at 768p. Renderforest's free plan includes limited H3 generations with no card required; paid plans remove the watermark and add commercial rights.
Does MiniMax H3 generate sound?
Yes — and this is its defining feature. H3 predicts video and audio together, producing 32kHz stereo covering dialogue, sound effects and ambience in the same pass. Because picture and sound come from one model, lip movement and impacts are synchronised rather than approximated.
How long can a MiniMax H3 video be?
Four to fifteen seconds per generation, in whole seconds, at 24 frames per second. For anything longer, generate several shots and stitch them in the Renderforest editor — you can cut H3 clips together with Veo 3.1 or Kling shots in the same timeline and export as one video.
What resolution does MiniMax H3 output?
Up to 2K, with a 1440-pixel short edge, in aspect ratios from 21:9 through to 9:16. The 2K is generated rather than upscaled afterwards, which is why on-screen text, logos and fine detail hold up better than in models that enlarge a lower-resolution render. Note that the open weights are limited to 768 pixels on the short edge.
MiniMax H3 vs Veo 3 — which should I use?
H3 for generated audio at 2K, multi-shot scenes and multilingual dialogue at a much lower cost per second. Veo 3.1 for maximum resolution and length — 4K with scene extension past a minute. Both run in Renderforest on the same plan, so run your prompt on each and compare rather than choosing in advance.
How is H3 different from Hailuo 2.3?
Hailuo 2.3 produced silent 768p or 1080p clips of fixed 6 or 10 seconds from a single source image, with no multi-shot. H3 produces 2K at any length from 4 to 15 seconds, with stereo audio generated alongside the picture, multiple shots from one prompt, and up to nine reference images plus video and audio. Prompts transfer with minimal changes.
Can I run MiniMax H3 locally?
If you are outside the US, EU, UK and South Korea, yes — the 33B weights are on Hugging Face, ComfyUI shipped native support on August 3, 2026, and AMD announced day-0 support on Instinct MI355X and MI300X. Two caveats: commercial use above $20M in annual revenue needs written authorisation, and because the 2K regeneration module was withheld, local output tops out at a 768-pixel short edge.
Can I use MiniMax H3 videos commercially?
Videos generated and edited in Renderforest on a paid plan come with commercial usage rights for ads, client work and monetised channels. Platforms including YouTube and Meta ask creators to label realistic AI-generated footage — worth doing regardless of which model produced it.
Will there be a Hailuo 4?
MiniMax has not announced one, and the release cadence has been quick — Hailuo 02 in 2025, Hailuo 2.3 that October, H3 nine months later. When the next version ships it lands in the same model menu here, which is the point of working on a multi-model platform rather than a single-model wrapper. We update this page the day anything changes.