What it is
MiniMax H3, also called Hailuo 3.0, was released on 31 July 2026, with its base weights on Hugging Face from early August. It takes text, images, video and audio as input and generates 4 to 15 second clips with 32 kHz stereo sound at 24 frames a second.
The open release is the 33B-parameter H3 transformer in two checkpoints (first/last frame and multi-reference, up to 9 images, 3 videos and 3 audio clips), producing video with a 768-pixel short side. It uses the full Qwen3-VL-32B model as its text and image encoder. The prompt-rewriting module (H3-Context-IR), which MiniMax strongly recommends, and the 2K regeneration step are API-only, and sparse attention is not yet in the open code.
The weights are under the MiniMax H3 Community License, which bars any use of the model or its outputs in the EU, UK, South Korea and the US without a separate licence from MiniMax, and requires written permission for products earning over $20 million a year.
Our review
As an API, H3 is the value pick among the leading video models. Artificial Analysis puts the 768p version at 1158 Elo on its with-audio text-to-video board, fourth and level with Seedance 2.5, at $4.80 a minute against Seedance's $34.12. The board's own confidence intervals overlap from third to eighth place, so the honest reading is that H3 sits in the leading pack, 17 points behind Wan 3.0, rather than below it.
The open-weights story is narrower than the label suggests. The licence defines the EU, UK, South Korea and the United States as excluded territories where the model and its outputs may not be used at all; anyone there has to contact MiniMax for a separate licence, and the card gives no process or timeline. Elsewhere, commercial use is allowed below $20 million in annual revenue, products must show "MiniMax H3" prominently in their interface, and outputs may not be used to train any other model. Military uses are banned outright.
Self-hosting is also heavier than the 33B headline implies. The repository holds about 500 GB of files, the encoder is the full 32B Qwen3-VL model, and the reference SGLang command spreads one job over four GPUs; the card publishes no VRAM figure or time per clip. The best-quality path still goes through MiniMax's servers, since the recommended prompt-rewriting module and the 2K regeneration step are hosted only.
For teams outside the excluded regions, H3 remains the one top-tier video model whose weights can be fine-tuned, and nothing else open comes close: Lightricks' LTX-2.5 Fast, the next open entry, scores 949. The 15-second clip ceiling is the practical limit, where Wan 3.0 runs to 30.
Where it earns its place
A performance-marketing agency producing 1,000 short ad variants a month
Generate 10-second product clips with sound from a packshot and a script
Call the MiniMax API with the first-frame checkpoint at 768p: 10,000 seconds costs about $800, against roughly $5,700 on Seedance 2.5 at its listed per-minute rate. Upscaling the winners to 2K adds $0.05 a second plus re-billed inputs.
An ML engineer at a studio in Singapore or the Gulf with a four-GPU node
Fine-tune a video model on the studio's own characters and style
Download the Ref2VA checkpoint and serve it with SGLang across four GPUs, as the card's example does. The licence permits this outside the excluded territories, but any fine-tuned model inherits the same licence and outputs cannot train a different model.
A product lead at a US or EU app company
Ship H3-powered video to customers
Use the hosted API or Runway rather than the weights: the community licence forbids using the model or its outputs in the US and EU without a separate deal with MiniMax. Budget $0.08 to $0.10 a second at 768p.
Before you commit
- Setup
- API: platform.minimax.io, billed per second of output. Weights: two BF16 checkpoints on Hugging Face (about 500 GB of repository files) with SGLang, vLLM, Diffusers and ComfyUI templates.
- Integrations
- MiniMax API (global and China endpoints), Runway API, Hailuo app; open-source serving through SGLang and vLLM, plus ComfyUI workflows for text-to-video and reference-to-video.
- Data and privacy
- Licence: no use of the model or outputs in the EU, UK, South Korea or US without a separate licence; written permission above $20 million annual revenue; "MiniMax H3" shown in the UI; no training other models on outputs. MiniMax's API retention and training terms were not published on the pages we could read.
- Cost at scale
- API: 1,000 ten-second clips a month is about $800 at 768p or $1,300 at 2K; images beyond the first five cost $0.04 each. Self-hosted: the card's reference job uses four GPUs, but GPU class and VRAM are not published.
- Limits
- 4 to 15 seconds per clip at 24 fps; open weights cap at 768p and use full attention only. Reference inputs are capped at 12 files, with each clip 2 to 15 seconds. Prepaid video packages do not yet cover H3.
What's good
- Fourth on Artificial Analysis's with-audio video board at $4.80 a minute
- Top open-weights video model by about 200 Elo
- Native 32 kHz stereo audio and up to 12 reference files
- Runs in SGLang, vLLM, Diffusers and ComfyUI
What's not
- Licence bars the model and its outputs in the US, EU, UK and South Korea
- Open weights stop at 768p; prompt rewriting and 2K are API-only
- 15-second clip limit
- No published VRAM figure; reference setup uses four GPUs
Best for
Teams that want leading video quality cheaply through an API, and teams outside the US, EU, UK and Korea that want to self-host or fine-tune a top video model.
Alternatives
MiniMax H3 (Hailuo 3.0): common questions
- Is MiniMax H3 (Hailuo 3.0) worth it?
- Near-frontier video at a seventh of Seedance's price, though the open weights are off limits in the four markets most buyers sit in. We rate it 4.3 out of 5.
- How much does MiniMax H3 (Hailuo 3.0) cost?
- MiniMax API: $0.08 per second at 768p, $0.13 at 2K ($4.80 per minute at 768p); 2K upscale of an existing clip $0.05 per second. Runway: $0.10 per second at 768p. Weights free under the MiniMax H3 Community License.
- What are the downsides of MiniMax H3 (Hailuo 3.0)?
- Licence bars the model and its outputs in the US, EU, UK and South Korea; open weights stop at 768p; prompt rewriting and 2K are API-only; 15-second clip limit; no published VRAM figure; reference setup uses four GPUs.
- Who is MiniMax H3 (Hailuo 3.0) best for?
- Teams that want leading video quality cheaply through an API, and teams outside the US, EU, UK and Korea that want to self-host or fine-tune a top video model.
- How much does MiniMax H3 (Hailuo 3.0) cost for a team or in production?
- API: 1,000 ten-second clips a month is about $800 at 768p or $1,300 at 2K; images beyond the first five cost $0.04 each. Self-hosted: the card's reference job uses four GPUs, but GPU class and VRAM are not published.
- Is MiniMax H3 (Hailuo 3.0) safe for company data?
- Licence: no use of the model or outputs in the EU, UK, South Korea or US without a separate licence; written permission above $20 million annual revenue; "MiniMax H3" shown in the UI; no training other models on outputs. MiniMax's API retention and training terms were not published on the pages we could read.
- What are the best MiniMax H3 (Hailuo 3.0) alternatives?
- The closest alternatives we have reviewed are Wan 3.0, Seedance 2.5 and FLUX 3.
Answers checked Oct 23, 2026.
Make MiniMax H3 (Hailuo 3.0)? Show this rating on your site
