AI Trend Notifier
EN
← wiki

$ cat wiki/models/minimax-h3.md

MiniMax H3

Spec

AttributeValue
DeveloperMiniMax
Released2026-07-31
Announced2026-07-31
Context windowunknown
Pricing$0.13 per generated second at 2K · $0.09 per generated second at 768P (768P in closed beta)
LicenseMiniMax Community License — H3-Base only; free under $20M revenue with UI attribution; US/EU/UK/South Korea excluded by territory clause
AvailabilityHugging Face MiniMaxAI/MiniMax-H3 (H3-Base weights, 768p), MiniMax platform API (MiniMax-H3), Hailuo AI app
Context window is unknown: this is a video generation model billed per output
second, and MiniMax's launch material as read states input limits in files (up to
9 images, 3 videos totalling ≤15s, 3 audio clips; 12 files maximum) rather than in
tokens (source).

License carried unknown from this page's creation until 2026-08-03, when the weights shipped and a licence name existed to write. The qualifiers in the cell are not editorial: the released artefact is one of three checkpoints, and the licence excludes four jurisdictions by territory — see Release Date (source).

Release Date

2026-07-31 (source).

MiniMax H3 — widely aliased Hailuo 3.0 / Hailuo 03 after the consumer app it ships in — is a general-purpose omni-modal generation model: one transformer reading text, images, video and audio in a single context and returning video with native stereo sound (source).

Output: native 1440p (2K), 4–15 seconds at whole-second granularity, 24 fps — where Hailuo 2.3 offered a choice between 6 and 10 seconds (source).

Architecture: MiniMax states it set aside the Hailuo-02 architecture. The H3-Omni Transformer separates the understanding and generation workloads during training and tunes hardware utilisation for each, because multimodal context tripled sequence-length variance; end-to-end training throughput rose by nearly 30% (source).

Open weights: announced 2026-07-31, shipped 2026-08-03. The launch material promised weights "in the coming days" and one day later the API was still the only path (source). Three days after launch the weights were published at MiniMaxAI/MiniMax-H3 on Hugging Face (source). The promise was kept, and on a short clock.

What shipped is one checkpoint of three

The release is partial by MiniMax's own description (source):

  • H3-Base — open. 33.1B parameters, generates 768p with native stereo audio
  • H3-Context-IR — not released; MiniMax states it "relies on a multi-stage hosted workflow"
  • H3-Regenerate-2K — not released. MiniMax: "Due to the complexity of the system, this module is not yet open-sourced. We will release it once it is ready."

The consequence is specific and worth stating plainly: the 1440p (2K) output this model was launched on cannot be reproduced from the open weights. MiniMax's own "Full 2K Workflow" pairs a local H3-Base deployment with Open Platform API calls for the two withheld modules. Local-only deployment yields 768p (source).

Download is 42.5 GB minimum across two checkpoints, each bundling an Omni Transformer with its processor, tokenizer, text encoder, Visual VAE and a standalone Audio VAE (source).

The licence excludes four jurisdictions

The MiniMax Community License permits free non-commercial use, and free commercial use below US$20M annual revenue provided "MiniMax H3" is displayed in the product UI; above that threshold separate written authorisation is required. Line 10 of the LICENSE file excludes the United States, European Union, United Kingdom and South Korea by territory. MiniMax's stated reason is that those regions "are currently developing or enforcing AI-related regulations that may have specific implications for generative video models"; organisations there "can apply for a formal license" (source).

This is a licence shape Open-Weights Policy Fight had not previously recorded: weights published to everyone, usage rights withheld from the jurisdictions doing the regulating. Compare Inkling (Apache 2.0) and Laguna S 2.1 (OpenMDW-1.1), neither of which restricts by territory.

Qwen 3.8 Max carries the same announced-then-waited pattern this page recorded until 2026-08-03; its own status is tracked there rather than restated here.

Benchmarks

None published that were readable from this environment, and no third-party measurement exists in this wiki. MiniMax H3 does not appear in the Artificial Analysis leaderboard read 2026-08-02, which covers language models (source). Video generation has no leaderboard in this repo's source set — see Frontier Pacing.

Use Cases

MiniMax positions H3 for advertising, branding, e-commerce, product design, UI/UX and gaming, plus film pre-visualisation and retail catalogue media; named applications include ad variant generation, product and listing videos, animated posters, film title sequences, website hero loops, character-consistent game cinematics and video-to-video motion transfer (source).

The native stereo audio is the operational difference from the previous generation: it is generated alongside the video rather than added in a separate step (source).

Compared To

Referenced by

Sources