Hailuo AI

Hailuo AI (海螺AI) is the consumer AI video generation service operated by MiniMax, a Shanghai-based foundation model developer that has traded on the Hong Kong Stock Exchange as MiniMax Group Inc. (stock code 00100) since January 2026.^[19] It produces short video clips from text prompts, still images, reference clips and reference audio, and is reached through the international site hailuoai.video, the Chinese-language site hailuoai.com, mobile applications, and MiniMax's developer API.^[2][26] MiniMax announced its first video model, Video-01, on August 31, 2024, and has shipped a new generation roughly every year since: Hailuo 02 in June 2025, Hailuo 2.3 in October 2025, and MiniMax H3 on July 31, 2026.^{[6][7][8][13]}

As of August 1, 2026 the current model is MiniMax H3, which generates 2K video with native stereo sound in clips of 4 to 15 seconds and accepts text, image, video and audio as context.^[1][2][3] In the Artificial Analysis Video Arena, which ranks systems by blind pairwise votes rather than by vendor benchmarks, H3 sat first on the video editing leaderboard and second on the text-to-video leaderboard when those pages were checked on August 1, 2026.^[15][17] MiniMax reported that its video models had been used to generate more than 600 million videos in total as of the end of 2025.^[19]

Overview

Hailuo AI is MiniMax's creative front end for video. The web product exposes video generation, image generation, audio, an agent mode and a shared asset workspace; the same models are sold to developers through MiniMax's Open Platform API.^[2][26] The Chinese-language site hailuoai.com carries the same H3 launch material under the name 海螺视频 (Hailuo Video), so the product is not restricted to overseas users the way some Chinese AI services are.^[26]

Generation modes have changed with each model generation. The Hailuo 01 family split capabilities across separate checkpoints: text-to-video, image-to-video, subject reference from a single portrait, a variant tuned for 2D illustration, and "Director" variants that accepted camera instructions in plain language.^[10][11][12] Hailuo 02 and 2.3 collapsed this to text-to-video and image-to-video.^[3] H3 replaces the task-specific split entirely with three documented modes: text-to-video, first-frame and last-frame image-to-video, and reference generation, in which up to nine images, three video clips and three audio clips are supplied as context and the relationship between that context and the desired output is described in words.^[2]

Naming

MiniMax's naming is easy to misread, and several of its product names collide:

Scroll sideways for more →

A language model release such as MiniMax M3 (June 2026) is not a video release, and MiniMax's version numbers do not run in parallel across modalities.^[6]

Development history

Hailuo 01 generation

MiniMax announced Video-01 on August 31, 2024, describing it as its first AI-native video generation model: 720p output at 25 frames per second, cinematic camera movement, and clips of up to six seconds, with 10-second output flagged for the next major version.^[13] A KR-Asia hands-on published on September 11, 2024 confirmed 1280x720 at 25 fps and six-second clips, and noted that image-to-video was still a planned feature at that point.^[14] Image-to-video was added later in 2024, and MiniMax's Video-01 page now documents both text-to-video and image-to-video modes.^[13]

Three specialised checkpoints followed in early 2025. S2V-01, announced on January 10, 2025, generated video from a single reference image while holding facial identity stable across camera angles and movement; MiniMax described it as the first model of its kind to do so without fine-tuning, and said posture, expression and lighting could be adjusted by text prompt.^[10] I2V-01-Live, announced on January 12, 2025, was tuned for animating 2D illustrations with stable, subtle expression changes.^[11] T2V-01-Director and I2V-01-Director added camera control: reduced movement randomness, preset camera configurations and improved prompt adherence, so pans, tracking shots and similar moves could be requested in plain language.^[12] MiniMax's platform release notes date the Director models to February 11, 2025, while the announcement post on its own site is dated March 3, 2025.^[6][12]

Hailuo 02

Hailuo 02 launched on June 18, 2025 with native 1080p output and a new architecture MiniMax called Noise-aware Compute Redistribution.^[7] MiniMax said the architecture raised training and inference efficiency by 2.5 times at a comparable parameter scale, which let it expand the model to three times the parameter count of its predecessor and quadruple the training data without raising prices.^[7] The company headlined instruction following and physical realism, and claimed that for highly intricate scenarios such as gymnastics Hailuo 02 was "currently the only model globally" capable of that performance, a claim it did not support with published measurements.^[7] In the same post MiniMax disclosed that Hailuo Video 01 had by then been used to generate over 370 million videos.^[7]

Independent corroboration arrived the same day: Artificial Analysis reported that Hailuo 02 (0616) entered its Image to Video Leaderboard at rank 2, ahead of Google Veo 3 without audio and behind ByteDance's Seedance 1.0, which was not then internationally available. Artificial Analysis noted that this made Hailuo 02 the highest-performing image-to-video model available to most users at that moment, and contrasted its $0.49 price for a 1080p six-second clip with Veo 3's $0.50 per second.^[18]

Hailuo 2.3

Hailuo 2.3 shipped on October 28, 2025 as an incremental improvement on the Hailuo 02 base, targeting three areas: physical action and body movement, stylisation (anime, illustration, ink wash painting and game CG), and live-action facial micro-expressions.^[8] MiniMax held pricing at Hailuo 02 levels and added a Hailuo 2.3 Fast variant that it said cut batch creation costs by up to 50 percent; MiniMax's own model listing shows the Fast variant as image-to-video only.^[3][8] The release went out simultaneously on the website, mobile apps and the API, and the Hailuo Video Agent was renamed the Media Agent at the same time.^[8] The online editor VEED became a day-one launch partner, exposing both Hailuo 2.3 and 2.3 Fast inside its AI Playground.^[23]

MiniMax H3

MiniMax announced H3 on July 31, 2026 and described it as a general-purpose multimodal generation model rather than a video model with bolted-on extras.^[1] H3 understands unified context across text, images, video and audio, and generates video with native stereo sound at up to 2K resolution and 15 seconds.^[1] MiniMax framed the design goal as removing task boundaries: image generation, audio generation and video generation had each fragmented into separate expert models for editing, subject reference, motion reference and so on, and H3 was pretrained to treat reference and editing relationships as natural-language descriptions instead of a fixed task set.^[1]

MiniMax said it deliberately abandoned the Hailuo 02 architecture for H3 because that design "would introduce unnecessary complexity for a model built around task generalization".^[1] On pricing it claimed that H3's per-second price at 2K is less than a third of mainstream models, and less than half at 768p compared with rivals' 720p.^[1] The company also said it planned to publish the model weights "in the coming days, subject to applicable laws and regulations", partly to widen hardware compatibility; no H3 weights had appeared in MiniMax's Hugging Face organisation as of August 1, 2026.^[1][25] A full technical report was promised but not yet published at that date.^[1] The weights followed on August 3, 2026, when MiniMax published the MiniMax H3 repository on Hugging Face and ModelScope under the MiniMax H3 Community License Agreement, a custom license dated August 2, 2026 whose licensed territory excludes the European Union, the United Kingdom, South Korea and the United States; the open release covers the 768p base model, while the H3-Context-IR preprocessing stage and the 2K regeneration stage remain API-only.^[28][29]

Models

Scroll sideways for more →

MiniMax's API documentation carries the Hailuo 02, 2.3 and 2.3 Fast entries under a "Legacy Models" heading as of August 1, 2026, with H3 as the sole current video model; the legacy endpoints remain callable.^[3]

The H3 endpoint is /v2/video_generation on api.minimax.io, with model identifier MiniMax-H3. Generation is asynchronous: a request returns a task_id that is polled until the response carries a download URL. Requests use a multimodal content array whose items are typed (text, image_url, video_url, audio_url) and labelled with a role such as first_frame, last_frame, reference_image, reference_video or reference_audio.^[2] Documented input limits are up to nine reference images, three reference video clips (2 to 15 seconds each, 15 seconds total) and three reference audio clips, capped at 12 files per request; image and video dimensions must fall between 256 and 5760 pixels with an aspect ratio between 2:5 and 5:2; the prompt limit is 7,000 characters and the request body limit is 64 MB.^[2] Aspect ratio is explicit for text-to-video and adaptive when a first frame image is supplied.^[2]

Video Agent and Media Agent

The Hailuo Video Agent entered beta on June 20, 2025, two days after Hailuo 02.^[9] Rather than a node-based workflow editor, it uses large language model tool invocation to drive a toolkit spanning ideation, asset sourcing, storyboarding, editing and voiceover, and exposes its step-by-step reasoning so a user can intervene at any stage.^[9] MiniMax set out three stages: prebuilt templates with one-click generation, then semi-customisable workflows where script, visuals and voiceover can each be edited, then a fully autonomous end-to-end agent.^[9]

On October 28, 2025 the Video Agent was renamed the Media Agent and extended to multi-modal creation, automatically selecting which models to call for a given brief and accepting uploaded images, video or audio for step-by-step work.^[8] The template-driven part of the system is also available as an API: a request supplies a template_id plus media and text inputs, and returns a finished video URL.^[27]

Architecture

MiniMax has published design notes rather than a full technical report for either recent generation, so the descriptions below are the company's own account.

Hailuo 02 introduced Noise-aware Compute Redistribution (NCR), which MiniMax describes as reallocating compute during training and inference instead of spending it uniformly, yielding a 2.5 times efficiency gain at a comparable parameter scale and thereby funding a threefold parameter increase and a fourfold increase in training data.^[7] MiniMax has not published the parameter counts themselves.

For H3, MiniMax abandoned the Hailuo 02 architecture and listed four components:^[1]

• Contextual Omni Representation, a captioning approach in which language describes not only the target video but the relationships between context items and the target, and among elements of the context. MiniMax says source material typically requires around 100,000 tokens of inference, distilled to roughly 4,000 tokens on average.
• H3-VAE, a rebuilt tokeniser whose higher compression ratio delivers what MiniMax calls a fourfold gain in effective sequence length, described as the enabling technology for native 2K output.
• H3-Omni Transformer, a training architecture that separates understanding and generation workloads and tunes hardware utilisation for each, which MiniMax says lifted end-to-end training throughput by nearly 30 percent.
• In-context Regeneration, in which the base model regenerates its own low-resolution output with the original multimodal context still in hand, in place of a conventional super-resolution module. MiniMax argues this recovers small text and fine detail that upscalers can only guess at.

MiniMax also states that all H3 audio output is native stereo and that voice, sound effects and music are modelled jointly rather than as separate domains, and that multi-shot structure is modelled natively.^[1]

Pricing and access

Prices and credit allowances on this service change frequently; the figures below were taken from MiniMax's own pages on August 1, 2026.

Consumer subscriptions

Hailuo AI runs on credits. Its subscription page lists five paid tiers plus pay-as-you-go credit purchases:^[20]

Scroll sideways for more →

MiniMax expresses the allowances in Hailuo 02 output: 1,000 credits is quoted as roughly 40 six-second 768p clips, 20 ten-second 768p clips or 12 six-second 1080p clips.^[20] Separately purchased credits expire after one year and do not confer subscriber benefits such as watermark removal or queue priority, and discounted introductory prices were being advertised alongside the list prices on the date checked.^[20] Watermark-free download is listed as a paid benefit, which implies watermarking on free output.^[20]

API pricing

Pay-as-you-go list prices on MiniMax's Open Platform:^[4]

Scroll sideways for more →

For H3, input audio is free, the first five input images are free and further images cost $0.04 each, and input video is billed by its duration at the output resolution's per-second rate.^[4] Prepaid video packages run from $1,000 for 3,760 video points to $6,000 for 26,780 points with request-rate limits attached; MiniMax's documentation states that these packages cover the Hailuo series only and that H3 is not yet supported.^[5]

Third-party availability

Hailuo models are resold through the generative media clouds, which is a usable availability signal because those listings are public:

• Replicate hosts minimax/hailuo-02, minimax/hailuo-02-fast, minimax/hailuo-2.3 and minimax/hailuo-2.3-fast. On August 1, 2026 the listed run counts were 436,100 for hailuo-02, 233,300 for hailuo-2.3-fast, 112,400 for hailuo-2.3 and 57,600 for hailuo-02-fast. H3 was not listed.^[21]
• fal.ai added Hailuo 2.3 on the day of its release, with Pro, Standard and Fast endpoints for both text-to-video and image-to-video.^[22]
• VEED exposes Hailuo 2.3 and 2.3 Fast inside its browser-based editor, having been a day-one launch partner.^[23]
• Hugging Face carried no Hailuo or H3 weights in MiniMax's organisation as of August 1, 2026; the organisation's public repositories were language models plus a set of video tokenisers.^[25] The H3 weights were published there on August 3, 2026 as MiniMaxAI/MiniMax-H3.^[28]

Comparison with competing systems

Vendor comparisons in this category are unreliable, and MiniMax's own launch material makes ranking claims without published measurements.^[1][7] The most usable independent yardstick is the Artificial Analysis Video Arena, which scores models with an Elo rating derived from blind pairwise votes on outputs generated from the same prompt. The tables below are the "current models, with audio" views as they stood on August 1, 2026; the "Range" column in the source indicates the statistical rank interval, so nearby models may be tied.^[15][16][17]

Text-to-video leaderboard, with audio:^[15]

Scroll sideways for more →

Artificial Analysis placed H3 and Gemini Omni Flash in a joint rank range of 1 to 2 on that board, meaning the three-point Elo gap is inside the confidence interval.^[15] On the image-to-video board with audio, H3 ranked third at 1,184 Elo behind Dreamina Seedance 2.0 720p (1,196) and Gemini Omni Flash (1,194), again within an overlapping range.^[16] On the video editing board with audio, H3 ranked first at 1,130 Elo ahead of Gemini Omni Flash at 1,122.^[17] Excluding audio, Artificial Analysis listed H3 second on both text-to-video (1,305 Elo) and image-to-video (1,351 Elo), behind Gemini Omni Flash in each case.^[15][16]

MiniMax's competitive position has moved a great deal. Hailuo 02 entered at rank 2 on the image-to-video board in June 2025;^[18] by the time H3 shipped, the arena had absorbed releases from ByteDance, KlingAI, Google, Alibaba, Skywork and others, and no Hailuo 01 or 02 era model appears in the with-audio current-models view, those models producing silent video; they still rank mid-table on the boards that exclude audio.^[15][16] Price is part of the comparison: at the leaderboard's normalised measure, H3 was the cheapest of the top three text-to-video systems apart from Gemini Omni Flash, and roughly a third the cost of Kling 3.0 Pro and Veo 3.1.^[15]

Claims that any one system is best overall should be treated carefully. Sora did not appear in the current-models arena views checked here, while Runway and Pika appear well below the leaders, Runway's Aleph 2.0 sitting sixth on the video editing board. Absence from a leaderboard, or a position on one, is not a quality judgement.

Applications

MiniMax positions H3 for advertising, branding, e-commerce, product design, user interface work and games, and cites instruction following, accurate text and brand rendering, and video-to-video motion transfer as the capabilities that make it usable for commercial work.^[1] Documented product-level uses include short-form social video, animation of static illustration through the Live variant of the Hailuo 01 family, character-consistent scenes built from a single reference portrait, and template-driven promotional clips through the Media Agent.^{[8][9][10][11]} MiniMax cited e-commerce advertising produced by beta testers around China's Double 11 shopping festival as an early Hailuo 2.3 use case.^[8]

Reception

Adoption figures come from MiniMax and have not been independently audited. The company said Hailuo Video 01 had been used to generate over 370 million videos by June 2025, and that its video models collectively passed 600 million videos by the end of 2025.^[7][19] In its 2025 annual results, MiniMax reported revenue from AI-native products of $53.1 million, up 143.4 percent year over year, attributing part of that to continued adoption and monetisation of Hailuo AI; more than 70 percent of total revenue came from international markets.^[19]

Independent evaluation has been consistently favourable at the top of the market without ever showing a clear lead. Artificial Analysis ranked Hailuo 02 second on image-to-video at launch in June 2025 and called it the highest-performing image-to-video model then available to most users, while noting it trailed a ByteDance model that was not internationally released.^[18] H3's August 2026 standing follows the same pattern: first on one of three boards, second or third on the others, and priced below most systems ranked near it.^[15][16][17] Distribution through Replicate, fal.ai and VEED, and the day-zero fal.ai launch for Hailuo 2.3, indicate steady demand from the developer and creator tooling market.^[21][22][23]

Limitations

• Clip length remains short. Hailuo 02 and 2.3 cap at 10 seconds (and at 6 seconds for 1080p), and H3 caps at 15 seconds.^[2][3]
• H3's 768p tier was in closed beta and required contacting sales as of August 1, 2026, so the cheaper price point was not generally available.^[4]
• Prepaid video packages did not cover H3 as of August 1, 2026, leaving H3 on pay-as-you-go billing only.^[5]
• MiniMax released the H3 weights on August 3, 2026, but the open release is partial: the H3-Context-IR and 2K regeneration stages remain API-only, and the license excludes the European Union, the United Kingdom, South Korea and the United States from its licensed territory.^[28][29]
• Native audio is new with H3; the earlier Hailuo models are presented and evaluated as silent video systems.^[1][15]
• MiniMax's own list of H3 shortcomings covers multimodal understanding, model size, and visual detail in some scenarios, with scaling and higher resolutions named as future work.^[1]
• Several headline claims in MiniMax's launch posts, including the assertion that Hailuo 02 was the only model globally capable of rendering complex gymnastics, are unaccompanied by published measurement and should be read as marketing.^[7]

Copyright litigation

On September 16, 2025, Disney, NBCUniversal and Warner Bros. Discovery jointly sued MiniMax in the United States District Court for the Central District of California, alleging "willful and brazen" copyright infringement and describing a "bootlegging business model" built on the studios' characters. The complaint centres on the Hailuo image and video service.^[24] The case is covered in more detail in the article on MiniMax; it forms part of a wider wave of AI and copyright litigation against generative media providers.

Timeline

Scroll sideways for more →

See also

• MiniMax
• MiniMax H3
• AI video generation
• Text-to-video generation
• Yan Junjie
• Artificial Analysis

References

1. ^MiniMax. "MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities." July 31, 2026. minimax.io/...minimax-h3
2. ^MiniMax API Docs. "Video Generation." Accessed August 1, 2026. platform.minimax.io/...video-generation
3. ^MiniMax API Docs. "Models." Accessed August 1, 2026. platform.minimax.io/...models-intro
4. ^MiniMax API Docs. "Pay as You Go." Accessed August 1, 2026. platform.minimax.io/...pricing-paygo
5. ^MiniMax API Docs. "Video Packages." Accessed August 1, 2026. platform.minimax.io/...pricing-video
6. ^MiniMax API Docs. "Release Notes: Models." Accessed August 1, 2026. platform.minimax.io/...models
7. ^MiniMax. "MiniMax Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency." June 18, 2025. minimax.io/...minimax-hailuo-02
8. ^MiniMax. "MiniMax Hailuo 2.3: A New Level of Complex Video Performance & Media Agent." October 28, 2025. minimax.io/...minimax-hailuo-23
9. ^MiniMax. "Introducing Hailuo Video Agent in Beta, Vibe Videoing with Zero-touch." June 20, 2025. minimax.io/...video-agent
10. ^MiniMax. "S2V-01: Subject Reference of Hailuo." January 10, 2025. minimax.io/...s2v-01-release
11. ^MiniMax. "I2V-01-live: smooth, vivid, and perfect for 2D illustrations." January 12, 2025. minimax.io/...i2v-01-live-release
12. ^MiniMax. "Hailuo AI Advances Cinematic Storytelling with T2V-01-Director and I2V-01-Director." March 3, 2025. minimax.io/...01-director
13. ^MiniMax. "MiniMax officially releases the Video-01 video generation model." August 31, 2024. minimax.io/...video-01
14. ^KR-Asia. "MiniMax's Hailuo AI text-to-video generator: How does it stack up?" September 11, 2024. kr-asia.com/...ideo-generator-how-does-it-stack-up
15. ^Artificial Analysis. "Text to Video Leaderboard." Accessed August 1, 2026. artificialanalysis.ai/...text-to-video
16. ^Artificial Analysis. "Image to Video Leaderboard." Accessed August 1, 2026. artificialanalysis.ai/...image-to-video
17. ^Artificial Analysis. "Video Editing Leaderboard." Accessed August 1, 2026. artificialanalysis.ai/...video-editing
18. ^Artificial Analysis (@ArtificialAnlys). "MiniMax's Hailuo 02 (0616) launches with rank #2 in the Artificial Analysis Image to Video Leaderboard." X, June 18, 2025. x.com/...1935335759701229857
19. ^MiniMax Group Inc. "MiniMax Announces Full Year 2025 Financial Results." PR Newswire, March 2, 2026. en.prnasia.com/...5-financial-results-523616.shtml
20. ^Hailuo AI. "Subscription Plans." Accessed August 1, 2026. hailuoai.video/subscribe
21. ^Replicate. "minimax" model listing. Accessed August 1, 2026. replicate.com/minimax
22. ^fal. "MiniMax Hailuo 2.3 is now available on fal." October 28, 2025. blog.fal.ai/...-hailuo-2-3-is-now-available-on-fal
23. ^MiniMax. "MiniMax and VEED: Introducing Hailuo-2.3 to Bring AI Video to Production Level." October 28, 2025. minimax.io/...eed-hailuo-23-for-pro-level-ai-video
24. ^TheWrap. "Disney, NBCUniversal, Warner Bros. Discovery Sue AI Firm MiniMax for Copyright Infringement." September 16, 2025. thewrap.com/...r-bros-discovery-minimax-ai-lawsuit
25. ^Hugging Face. Model listing for the MiniMaxAI organisation. Accessed August 1, 2026. huggingface.co/...models
26. ^Hailuo AI. Home page. Accessed August 1, 2026. hailuoai.video
27. ^MiniMax API Docs. "Video Generation with Templates Guide." Accessed August 1, 2026. platform.minimax.io/...video-agent
28. ^MiniMax. "Open General Intelligence: MiniMax H3 Is Now Open Source." August 3, 2026. minimax.io/...minimax-h3-open-source
29. ^MiniMax (Nanonoble Pte. Ltd.). "MiniMax H3 Community License Agreement." August 2, 2026. huggingface.co/...LICENSE