Hailuo AI
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v7 · 3,349 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v7 · 3,349 words
Add missing citations, update stale details, or suggest a clearer explanation.
Hailuo AI is an AI video generation platform developed by MiniMax, a Shanghai-based artificial intelligence company, launched in September 2024. It lets users create short video clips from text prompts, still images, and subject reference images using a suite of specialized generative AI video models.[14] Hailuo AI gained rapid international traction for producing fluid, lifelike human motion and cinematic visual quality, placing it among the top-tier video generation tools worldwide. By June 2025, the flagship Hailuo 02 model ranked #2 on the Artificial Analysis Image-to-Video Leaderboard, beating Google's Veo 3 (without audio) and trailing only ByteDance's Seedance 1.0.[8] As of the end of 2025, MiniMax's video models had helped creators worldwide generate more than 600 million videos in total.[12]
Hailuo AI is the consumer-facing video generation brand of MiniMax. It is accessible through its web interface at hailuoai.video, through mobile applications on iOS and Android, and via API endpoints for developers. The platform supports over 70 languages and serves a global user base. As of January 2025, Hailuo AI ranked 12th globally in user traffic among AI tools, and by October 2025 the overseas version reported roughly 11.73 million monthly visits.
The platform offers several core generation modes:
MiniMax Group Inc. was founded in December 2021 by Yan Junjie and Zhou Yucong, both former researchers at SenseTime. The company is headquartered in Shanghai, China, and develops multimodal AI models along with consumer-facing applications.
Yan Junjie, the CEO, was born in 1989 in Henan province. He earned a PhD from the Institute of Automation at the Chinese Academy of Sciences and later completed postdoctoral research at Tsinghua University, focusing on deep learning and computer vision. Before founding MiniMax, Yan spent over six years at SenseTime, where he rose to the position of vice president. His interest in large-scale AI systems grew after watching OpenAI Five defeat professional Dota 2 teams in 2019, which prompted his shift from computer vision to natural language processing.[13]
Zhou Yucong, the co-founder, previously led algorithms R&D at SenseTime and joined MiniMax in early 2022 to oversee research and development of visual models.
By the end of 2025, MiniMax reported that it had cumulatively served more than 236 million users across over 200 countries and regions, along with roughly 214,000 enterprise clients and developers in more than 100 countries.[12]
MiniMax raised approximately $850 million across multiple funding rounds before going public. In March 2024, Alibaba Group led a $600 million Series A round that valued the company at $2.5 billion. Other investors included Tencent Holdings, Hillhouse Investment, HongShan (formerly Sequoia Capital China), IDG Capital, and miHoYo.
On January 9, 2026, MiniMax listed on the Hong Kong Stock Exchange in what became the largest IPO among AI foundation model companies at the time.[9] The company priced shares at HK$165 (the top of the indicated range), raising approximately HK$5.54 billion (roughly $619 million).[11] Shares surged 109% on the first day of trading, closing at HK$345 and pushing the market capitalization above HK$76.3 billion (about $9.8 billion).[10] The retail tranche was oversubscribed by 1,837 times. Cornerstone investors included Alibaba, the Abu Dhabi Investment Authority (ADIA), Aspex, Eastspring, and Mirae Asset. Following the IPO, Yan Junjie's net worth reached an estimated $3.2 billion.[13]
For full-year 2025, MiniMax reported total revenue of $79.0 million, a 158.9% increase year over year. Revenue from AI-native products (including Hailuo AI, Talkie, and Xingye) reached $53.1 million, up 143.4% from $21.8 million in 2024, while revenue from the Open Platform and enterprise AI services contributed $26.0 million. Over 70% of total revenue came from international markets.[12]
Beyond Hailuo AI, MiniMax operates several consumer products:
Hailuo AI's first model, Video-01, launched in early September 2024 as a text-to-video generator. It produced 1280x720 resolution videos at 25 frames per second with a maximum duration of six seconds. Generation time was roughly two minutes per clip. At launch, the platform offered free access to all users, which helped it quickly build a following among content creators.[14]
Video-01 stood out for generating human motions that appeared notably more fluid and lifelike than those of competing tools at the time, including offerings from Runway and Luma AI. The model supported only text-to-video at launch.[14]
On October 8, 2024, MiniMax expanded Hailuo AI with the I2V-01 model, introducing image-to-video generation. This allowed users to upload a still image and transform it into a short animated clip. The initial I2V output was limited to approximately four seconds.[15]
Released on December 3, 2024, I2V-01-Live was designed specifically for animating 2D illustrations. It was among the first image-to-video models purposefully trained for Live2D-style use cases. The model excelled at converting manga, anime, digital portraits, and other illustrated artwork into smooth animations with expressive character movements. It supported resolutions from 512p to 1080p, durations up to 10 seconds (6 seconds at 1080p), and aspect ratios from 2:5 to 5:2.[6]
The Director variants, announced in early 2025 and formally released on March 3, 2025, introduced precise camera control capabilities through natural language commands. T2V-01-Director converts text prompts to video with enhanced cinematic framing, while I2V-01-Director does the same starting from still images.[4]
Key improvements in the Director models included reduced movement randomness, pre-configured cinematic camera settings, and stronger prompt adherence. These models allowed creators to specify camera pans, zooms, tracking shots, and other cinematographic techniques using plain text instructions rather than technical parameters.[4]
Released on January 10, 2025, S2V-01 addressed one of the most persistent challenges in AI video generation: maintaining consistent character identity across different scenes. The model generates videos from a single reference image while preserving facial features and identity regardless of camera angles, lighting changes, or character movements.[3]
S2V-01 allows adjustments to posture, expressions, and lighting through text-based prompts. MiniMax described it as the first model of its kind to ensure character consistency in dynamic video generation without requiring fine-tuning. The development roadmap includes planned support for multi-subject references, object references, and multi-layered scenes.[3]
Hailuo 02 launched on June 18, 2025, representing a major architectural overhaul. It introduced the Noise-aware Compute Redistribution (NCR) architecture, a novel framework that dynamically redistributes computational resources based on scene complexity during both training and inference.[1]
The NCR approach allocates more processing power to harder-to-render frames and visual elements, improving efficiency by 2.5 times compared to the previous generation at comparable parameter scales. This efficiency gain allowed MiniMax to scale the model to 3 times the parameter count of its predecessor while keeping costs manageable for users. Training data volume was also expanded by 4 times, with improvements in both quality and diversity.[1]
The model incorporated a transformer-based backbone tailored for video diffusion, along with a Mixture-of-Experts (MoE) component that enables specialized sub-networks to handle different visual and motion subtasks.[1]
MiniMax highlighted the model's physical realism as a headline capability, describing Hailuo 02 as delivering "Extreme Physics Mastery" and noting that for highly intricate scenarios such as gymnastics it was "currently the only model globally" capable of such performance.[1]
Hailuo 02 output options include:
Upon launch, Hailuo 02 achieved the #2 position on the Artificial Analysis Image-to-Video Leaderboard, behind only ByteDance's Seedance 1.0 (which was not yet internationally available at the time) and ahead of Google's Veo 3 (without audio).[8] The model received a score of 92.1 on the Artificial Analysis benchmark compared to Veo 3's 87.3, with a physics simulation rating of 94 out of 100.[1]
Released on October 28, 2025, Hailuo 2.3 built upon the Hailuo 02 architecture with targeted improvements in three areas:[2]
Hailuo 2.3 also introduced a Fast variant that generates videos more quickly at reduced cost, cutting batch creation expenses by up to 50%. The standard Hailuo 2.3 maintained the same pricing as Hailuo 02.[2]
Alongside Hailuo 2.3, MiniMax rebranded the Hailuo Video Agent as the Media Agent, a broader multi-modal creation tool that automatically selects the appropriate models based on user input.[2]
| Model | Release Date | Generation Mode | Max Resolution | Max Duration | Key Feature |
|---|---|---|---|---|---|
| Video-01 | September 2024 | Text-to-Video | 720p (1280x720) | 6 seconds | First Hailuo model; fluid human motion |
| I2V-01 | October 2024 | Image-to-Video | 720p | 4 seconds | Still image animation |
| I2V-01-Live | December 2024 | Image-to-Video | 1080p | 10 seconds | Optimized for 2D/anime illustration |
| S2V-01 | January 2025 | Subject-to-Video | 720p | 6 seconds | Character consistency from single reference |
| T2V-01-Director | March 2025 | Text-to-Video | 720p | 6 seconds | Natural language camera control |
| I2V-01-Director | March 2025 | Image-to-Video | 720p | 6 seconds | Camera control from still images |
| Hailuo 02 | June 2025 | T2V, I2V | 1080p | 10 seconds | NCR architecture; 3x parameters |
| Hailuo 2.3 | October 2025 | T2V, I2V | 1080p | 10 seconds | Enhanced motion, stylization, micro-expressions |
| Hailuo 2.3 Fast | October 2025 | T2V, I2V | 1080p | 10 seconds | 50% cost reduction for batch creation |
The Hailuo Video Agent entered beta on June 20, 2025, two days after the Hailuo 02 model launch. It is designed to handle the complete video creation pipeline through a single natural language prompt, replacing the need for manual storyboarding, editing, and post-production.[5]
The agent uses large language model-powered tool invocation to orchestrate a full-stack toolkit covering ideation, asset sourcing, storyboarding, editing, and voiceover generation. Users can view the agent's step-by-step reasoning process in real time, offering transparency into its creative decisions and allowing manual overrides at any stage.[5]
MiniMax outlined a three-stage development plan for the Video Agent:[5]
In October 2025, the Video Agent was rebranded as the Media Agent alongside the Hailuo 2.3 release. The Media Agent expanded capabilities to support comprehensive multi-modal creation, automatically matching the appropriate models to the content type being produced.[2]
The NCR architecture, introduced with Hailuo 02, represents the core technical innovation behind the platform's second-generation models. Traditional video generation models distribute computational resources uniformly across all frames and visual elements. NCR instead analyzes the "noise" or complexity of each frame segment and dynamically reallocates compute power, directing more resources toward visually complex portions such as fast motion, intricate textures, or difficult lighting conditions.[1]
This approach achieves a 2.5x improvement in training and inference efficiency, enabling MiniMax to train a model with 3x the parameter count of Video-01 without proportionally increasing costs.[1]
Hailuo 02 and subsequent models incorporate a Mixture-of-Experts (MoE) component within the transformer backbone. This allows specialized sub-networks to activate selectively depending on the visual or motion subtask at hand. For example, separate expert modules may handle human body dynamics, environmental rendering, and lighting simulation, improving output quality without requiring all parameters to be active for every generation step.[1]
Like most leading video generation systems, Hailuo AI models are built on a diffusion model framework. The generation process begins with noise and iteratively refines it into coherent video frames through a learned denoising process. The transformer-based backbone processes temporal and spatial information jointly, allowing the model to maintain consistency across frames.
Hailuo AI operates on a credit-based system. A standard 720p, six-second video generation typically costs 30 to 50 credits depending on settings and model version. The platform offers the following subscription tiers:[16]
| Plan | Monthly Price (USD) | Monthly Credits | Key Features |
|---|---|---|---|
| Free | $0 | Daily bonus credits | 720p exports, watermark, queue wait times |
| Standard | $14.99 | 1,000 | Queue priority, watermark-free, Hailuo 02 access |
| Pro | $54.99 | 4,500 | Accelerated generation, all models |
| Master | $119.99 | 10,000 | Highest priority, all features |
Note: Pricing tiers have been updated over time. An earlier Unlimited Plan ($94.99/month) was available for purchase before June 18, 2025, but it only supported the Hailuo 01 model.
New users receive a one-time welcome package of free credits that expire three days after registration. Additional credits can be purchased at approximately $1 per 70 credits.[16]
For developers accessing Hailuo models through the API, pricing is structured per second of generated video:
| Resolution | Approximate Cost per Second |
|---|---|
| 512p | ~$0.017 |
| 768p | ~$0.045 |
| 1080p | ~$0.082 |
API-level pricing translates to roughly $2.80 to $4.90 per minute of generated video, depending on resolution and model variant. Third-party platforms such as fal.ai and Replicate also offer access to Hailuo models at competitive rates.
Hailuo AI competes in a rapidly evolving video generation market alongside several well-funded platforms:
| Platform | Developer | Notable Strength |
|---|---|---|
| Sora | OpenAI | Photorealistic rendering, narrative coherence |
| Runway Gen-4.5 | Runway | Flexible experimentation, stylized creativity |
| Kling 3.0 | Kuaishou | Speed, high-volume social content |
| Veo 3 | Google DeepMind | Cinematic camera work, 4K resolution support |
| Seedance 2.0 | ByteDance | Top-ranked on Artificial Analysis leaderboards |
| Pika | Pika Labs | Rapid iteration, consumer accessibility |
| PixVerse | PixVerse | Strong visual quality, consistent updates |
The Artificial Analysis Video Arena, which ranks models based on blind user votes, shows the following landscape for text-to-video (without audio):[7]
| Rank | Model | Elo Score |
|---|---|---|
| 1 | Dreamina Seedance 2.0 (ByteDance) | 1,269 |
| 2 | Kling 3.0 1080p Pro (KlingAI) | 1,248 |
| 3 | SkyReels V4 (Skywork AI) | 1,247 |
| 7 | Runway Gen-4.5 | 1,226 |
| 10 | Veo 3 (Google) | 1,221 |
| 24 | Hailuo 2.3 (MiniMax) | 1,180 |
| 27 | Hailuo 02 Standard | 1,176 |
It is worth noting that leaderboard rankings fluctuate as new models are released and existing models receive updates. When Hailuo 02 first launched in June 2025, it achieved the #2 position on the Artificial Analysis Image-to-Video Leaderboard.[8] The competitive field has since intensified with major releases from ByteDance, Kuaishou, Google, and others in late 2025 and early 2026.
Hailuo AI's strengths remain strongest in human-centered content. The platform excels at rendering body movement, micro-expressions, and physical stability, making it particularly well-suited for character acting, emotional short films, and human-centered advertising. Its cost efficiency and physics-driven realism also give it an edge in markets where budget constraints and creative flexibility are primary considerations.
Unlike some Chinese AI platforms that remain restricted to domestic markets, Hailuo AI has pursued an international-first strategy. The platform is accessible worldwide through hailuoai.video and supports over 70 languages. MiniMax reported that over 70% of its 2025 revenue came from international markets, underscoring the success of this approach.[12]
Hailuo AI is available through multiple channels:
The platform has found particular traction among TikTok and YouTube content creators seeking affordable tools for short-form video production. MiniMax's partnership with VEED, announced alongside Hailuo 2.3, aimed to bring AI video capabilities to professional production workflows.[2]
Hailuo AI serves a wide range of creative and commercial applications:
| Date | Event |
|---|---|
| December 2021 | MiniMax founded by Yan Junjie and Zhou Yucong |
| March 2024 | Alibaba leads $600 million funding round at $2.5 billion valuation |
| September 2024 | Video-01 launches as Hailuo AI's first text-to-video model |
| October 2024 | I2V-01 adds image-to-video generation |
| December 2024 | I2V-01-Live released for 2D illustration animation |
| January 2025 | S2V-01 introduces subject-consistent video generation |
| March 2025 | T2V-01-Director and I2V-01-Director add camera control |
| June 18, 2025 | Hailuo 02 launches with NCR architecture; ranks #2 on Artificial Analysis I2V Leaderboard |
| June 20, 2025 | Video Agent enters beta |
| October 28, 2025 | Hailuo 2.3 released; Video Agent rebranded as Media Agent |
| January 9, 2026 | MiniMax IPO on Hong Kong Stock Exchange raises ~$619 million |
| March 2026 | MiniMax reports 2025 revenue of $79.0 million (158.9% YoY growth) |