Hailuo video models (H3) by MiniMax: status and timeline

Hailuo video models (H3)

Make cinematic, physically believable video with synchronized sound cheap enough for everyday creators and studios, competing with Kling, Seedance and Veo.

ScalingStage 5 of 5

H3 (Jul 2026) makes 2K clips up to 15 s with native audio; weights out, but not licensed in US, EU, UK, Korea.

Updated 3 Aug 2026Checked 11 Oct0 updates this week

Milestones

No announced next step

Hailuo Video 01 launchesAug 2024Complete.
Hailuo 02: native 1080p, 3x parameters18 Jun 2025Complete.
Hailuo 2.3 and 2.3 Fast28 Oct 2025Complete.
H3 launches: 2K video with native stereo audio31 Jul 2026Complete.
H3 weights published on Hugging Face3 Aug 2026Complete.

Most important updates

Current obstacles

Restricted licence for open weightsThe H3 licence excludes the US, EU, UK and South Korea, and firms above $20M revenue need permission, limiting Western adoption.
Crowded field of near-equalsKling, Seedance, Veo and Wan release new versions every few months, so any lead in quality or price is short-lived.
Slow generationA five-second 2K clip takes about three minutes to render on one API host, and a 15-second clip much longer, too slow for interactive editing.

Physics limits

Compute grows faster than clip lengthA video model attends across every frame at once; doubling length or resolution more than doubles the work, which is why clips stay around 10 to 15 seconds.
Learned physics is only statisticalThe model copies how things usually look in its training videos rather than simulating forces, so unusual motions, collisions and hands still go wrong.
Consistency across shotsEach clip is generated separately, so keeping a face, outfit or room identical across many shots depends on reference inputs and still drifts.

How it works

4 parts

Tokenizer

Compressing video into tokens

A new video encoder squeezes frames into fewer tokens, which MiniMax says gives four times the usable sequence length, so longer 2K clips fit.

Generator

Picture and sound together

One transformer generates the video and its stereo soundtrack jointly, so lips, footsteps and ambience line up without a separate audio model.

Upscaling

Draft low, then regenerate

The model first makes a low-resolution clip, then regenerates its own output at 2K, which is cheaper than generating full detail from the start.

References

Locking a character

Up to nine images, three videos and three audio clips pin down a person's look and voice, so the same character holds across separate shots.

Spec sheet

vs Hailuo 2.3

Spec	Hailuo video models (H3)	Hailuo 2.3	Change
Max resolution	2K^R (reported)	1080p (Hailuo 2.3)	—no comparable change
Clip length	4 to 15 s^R (reported)	6 s at 1080p, 10 s at 768p	▲ 9 s
Audio	Native stereo, same pass^R (reported)	None	—no comparable change
Base model size	33B, dense^R (reported)	—	—no comparable change
Reference inputs	9 images, 3 videos, 3 audio clips^R (reported)	—	—no comparable change
Weights, smallest quantized version	42.5 GB^R (reported)	—	—no comparable change

^R reported by the company

Update log

4 updates

About MiniMax

The team behind Hailuo video models (H3)

MiniMax

M3 428B MoE model, open weights, Hailuo AI, Talkie

Shanghai-based AI company founded in December 2021 by computer vision researchers from SenseTime. Develops multimodal models and consumer apps (Talkie, Hailuo AI). Listed on Hong Kong Stock Exchange in January 2026.

M3 (June 2026): about 428B parameters with 23B active per token and a 1M-token context; weights on Hugging Face.
M3 scores 80.5% on SWE-Bench Verified, up from M2.5's 80.2%.
M2 and M2.5 weights are MIT-licensed; M3 uses a MiniMax community licence.

minimaxi.com (opens minimaxi.com)Company profile →

Founded: 20215 yrs
Headquarters: China
Status: Public
Valuation: $14Bmarket capJan 2026
Last round: IPO · $620MJan 2026
Works in: AIFrontier models
Coverage: 2 programs · 8 updateslatest 3 Aug 2026checked 11 Oct
Partners: Hugging Face (opens huggingface.co)Hong Kong Stock Exchange (opens hkex.com.hk)
People: Yan Junjie (opens en.wikipedia.org)Founder & CEO

Spec

Hailuo video models (H3)

Hailuo 2.3

Change

Max resolution

2K^R (reported)

1080p (Hailuo 2.3)

—no comparable change

Clip length

4 to 15 s^R (reported)

6 s at 1080p, 10 s at 768p

▲ 9 s

Audio

Native stereo, same pass^R (reported)

None

—no comparable change

Base model size

33B, dense^R (reported)

—

—no comparable change

Reference inputs

9 images, 3 videos, 3 audio clips^R (reported)

—

—no comparable change

Weights, smallest quantized version

42.5 GB^R (reported)

—

—no comparable change

Hailuo video models (H3)

Milestones

Most important updates

Current obstacles

Physics limits

How it works

Compressing video into tokens

Picture and sound together

Draft low, then regenerate

Locking a character

Spec sheet

Update log

Mon 3 Aug

Fri 31 Jul

Tue 28 Oct 2025

Wed 18 Jun 2025

About MiniMax

Current obstacles

Physics limits

How it works

Compressing video into tokens

Picture and sound together

Draft low, then regenerate

Locking a character

Spec sheet

Update log

Mon 3 Aug

Fri 31 Jul

Tue 28 Oct 2025

Wed 18 Jun 2025

About MiniMax