MiniMax drops Hailuo 02 video model with extreme physics

If you have tried using AI to generate high-speed motion, you probably know the pain: human joints twisting backwards, sudden extra limbs, or bodies melting like rubber.

MiniMax just released Hailuo 02, its next-generation video generation model aimed directly at fixing those nightmare physics glitches.

According to the company, creators have already generated over 370 million videos using the previous Hailuo 01 since its quiet debut late last August. An early build of Hailuo 02 had already climbed to second place globally on the Artificial Analysis Video Arena leaderboard.

The updated model is now available across Hailuo's web platform, mobile app, and API.

A new architecture that triples model size

AI video models are notoriously expensive to train and run, largely because compute isn't always spent where it matters most during generation.

MiniMax built a new framework called Noise-aware Compute Redistribution (NCR). The core idea is simple: dynamically route compute based on the noise level across generation steps.

At a comparable scale, NCR boosts training and inference efficiency by 2.5 times. That extra breathing room allowed MiniMax to scale Hailuo 02's parameters to three times that of its predecessor, while multiplying training data by four.

The company claims this overhaul vastly improves prompt adherence and extreme physics handling. MiniMax even called it the only model worldwide capable of handling ultra-complex physical scenarios like gymnastics without breaking apart.

To showcase this, three artists collaborated over 1.5 days to create a full video using 6-to-10-second raw clips generated by the model.

In terms of output, Hailuo 02 delivers native 1080p resolution. Users can currently choose between three tiers: 768p at 6 seconds, 768p at 10 seconds, and 1080p at 6 seconds.

Why it matters

For digital creators and commercial animators, Hailuo 02 offers a desperately needed upgrade: native 1080p output and realistic human physics. Most AI video models shine at slow pans and talking heads, but completely break down when characters perform rapid acrobatic moves. If Hailuo 02 delivers stable motion mechanics consistently, it earns an immediate spot in real production pipelines.

On the industry side, MiniMax's engineering trade-off is remarkably smart. Rather than brute-forcing massive parameter counts and driving API costs through the roof, they used the NCR architecture to buy back a 2.5-fold efficiency gain first, then reinvested that budget into a 3x larger model. This keeps inference costs low enough for mass adoption while leveling up visual fidelity.

That said, the challenge remains consistency. While MiniMax promised upcoming upgrades in generation speed, alignment stability, and capabilities beyond basic text-to-video and image-to-video, users will still have to assemble short clips manually. Until AI can maintain long-context multi-shot continuity, Hailuo 02 is a powerful b-roll factory rather than a turnkey director.