MiniMax is putting price and performance at the center of its new AI video release. Hailuo 02, the company's second-generation video AI model, arrives with higher resolution options, broader generation capabilities, and benchmark results that place it ahead of Google's much-discussed Veo 3 in one user-rated comparison.
The launch matters because AI video tools are quickly moving from demos to repeatable production workflows. For creators and API users, the central question is no longer only whether a model can generate an impressive clip. It is also whether the model can handle complex prompts, produce useful resolution options, and do so at a cost that supports repeated use.
What MiniMax changed in Hailuo 02
Hailuo 02 is built around an architecture MiniMax calls Noise-aware Compute Redistribution, or NCR. According to MiniMax, NCR improves training and inference efficiency by a factor of 2.5.
The basic idea is that long video sequences are handled differently depending on the training stage. When artificial noise is heavily present early in training, videos are compressed as much as possible. Later, when the training videos are clearer, the model processes them at full resolution.
MiniMax also says Hailuo 02 is larger and trained more broadly than the previous version. The new model has three times more parameters and four times more training data than its predecessor, along with improvements in data quality and diversity. The company has not disclosed exact parameter counts or dataset sizes for Hailuo 02.
Those changes are presented as practical upgrades, not just technical scale. According to MiniMax, Hailuo 02 is better at following complex prompts and simulating physical processes. The company says it is currently the only model able to accurately generate intricate scenes like gymnastics routines.
Resolution, duration, and access options
Hailuo 02 is available in three variants: 768p for six seconds, 768p for ten seconds, and 1080p for six seconds. That is a step up from the previous model, which was limited to 720p, six-second videos at 25 fps.
The model can be used through a web interface, a mobile app, or an API. That range of access points makes the release relevant to different groups: individual creators testing prompts, mobile users generating clips directly, and developers adding AI video generation to products or workflows.
For API users, the price difference is one of the clearest parts of the release. Generating a six-second 768p video costs $0.28, while a 1080p version costs $0.49. By comparison, producing an eight-second 1080p video with Google Veo 3 can cost around $3, depending on the plan.
Those prices are not identical product comparisons because the durations differ. Still, the figures show why Hailuo 02 is likely to draw attention from users who need to generate many clips, iterate on prompts, or test image-to-video and text-to-video ideas before selecting final outputs.
How Hailuo 02 performed against rivals
In the Artificial Analysis Video Arena benchmark, users rate videos from competing AI models. Hailuo 02 finished second in the image-to-video category.
It placed behind Bytedance's Seedance and ahead of Google Veo 3. The source comparison notes an important limitation: this version of Veo 3 does not support audio, which is described as a major part of its appeal.
That caveat matters because AI video models are not judged on one capability alone. Visual quality, prompt adherence, motion, physical realism, duration, resolution, audio support, speed, stability, and price can all affect whether a tool fits a specific use case. In this benchmark, Hailuo 02's image-to-video results were strong enough to outrank Veo 3, but the comparison does not erase the broader differences between platforms.
MiniMax also has evidence of broad user interest in the Hailuo platform. Since its demo launch in August last year, people have created over 3.7 billion videos using Hailuo, according to MiniMax. The company describes the initial rollout as very random, but says it quickly attracted widespread attention from creators worldwide.
What still remains unclear
Hailuo 02 arrives during MiniMax Week, a five-day event in which the Chinese startup also unveiled MiniMax-M1, an open-source language model with parameter counts and a technical paper. By contrast, technical details about Hailuo 02's training architecture remain undisclosed.
MiniMax says it is working to improve generation speed and stability. It also plans to add new features beyond the current text-to-video and image-to-video options.
That future work is important because competing platforms already offer capabilities that Hailuo 02 does not yet match. Runway, for example, already offers more advanced capabilities such as tracking shots.
For now, Hailuo 02's pitch is clear: stronger user benchmark performance, higher available resolution than the previous Hailuo model, and API pricing that is much lower than the Google Veo 3 comparison given. The remaining question is how quickly MiniMax can expand features and disclose enough technical detail to satisfy users who want both practical results and a clearer view of how the system works.