# MiniMax

> Source: https://aiwiki.ai/wiki/minimax
> Summary: MiniMax (稀宇科技; listed as MiniMax Group Inc.) is a publicly traded artificial intelligence (AI) company headquartered in Shanghai, China that develops large language models, multimodal AI systems, and consumer-facing AI applications spanning text, video, speech, and music generation.
> Updated: 2026-09-23
> Fact-checked: 2026-09-23
> Categories: AI Companies, Chinese AI, Generative AI, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "MiniMax." aiwiki.ai, 23 Sept 2026. https://aiwiki.ai/wiki/minimax
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| MiniMax |
| --- |
| MiniMax Group Inc. (稀宇科技) |
| ![MiniMax logo](https://qqcb8dyk5bp2il4c.public.blob.vercel-storage.com/wiki/images/200px-minimax_logo1.png) |
| Type | Public (HKEX: 0100, short name MINIMAX-W)[16] |
| Industry | [Artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) |
| Founded | Early 2022 (holding company incorporated in the Cayman Islands on June 30, 2021)[4][7] |
| Founders | [Yan Junjie](https://aiwiki.ai/wiki/yan_junjie) (闫俊杰)
Yun Yeyi (贠烨祎)[4] |
| Headquarters | Xuhui District, Shanghai, China[4] |
| Key people | Yan Junjie (chairman, CEO and CTO)
Yun Yeyi (executive director and COO)[4][22] |
| Main operating subsidiaries | Shanghai Xiyu Jizhi Technology Co., Ltd. (上海稀宇极智科技有限公司); Shanghai Xiyu Technology Co., Ltd. (上海稀宇科技有限公司); Nanonoble Pte. Ltd. (Singapore)[4] |
| Listing entity | MiniMax Group Inc. (Cayman Islands, weighted voting rights) |
| Products | Talkie
Xingye (星野)
[Hailuo AI](https://aiwiki.ai/wiki/hailuo) (海螺AI)
[MiniMax Code](https://aiwiki.ai/wiki/minimax_code) (formerly MiniMax Agent)
MiniMax Open Platform (API)
abab series
MiniMax-01 series
[MiniMax-M1](https://aiwiki.ai/wiki/minimax_m1), [MiniMax-M2](https://aiwiki.ai/wiki/minimax_m2) series, [MiniMax M3](https://aiwiki.ai/wiki/minimax_m3)
Hailuo video models, [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3)
Speech and Music models |
| Revenue | US$79.0 million (2025); US$116.6 million (first half of 2026)[1][22] |
| Market value | About US$13.7 billion at the close of its first trading day (January 9, 2026)[3] |
| Employees | 568 full-time (June 30, 2026)[22] |
| Website | [minimaxi.com](https://minimaxi.com) |

**MiniMax** (稀宇科技; listed as MiniMax Group Inc.) is a publicly traded [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) (AI) company headquartered in Shanghai, China that develops [large language models](https://aiwiki.ai/wiki/large_language_model), multimodal AI systems, and consumer-facing AI applications spanning text, video, speech, and music generation. The company says it was founded in early 2022; its Cayman Islands holding company was incorporated on June 30, 2021, and its prospectus names chairman and CEO Yan Junjie and chief operating officer Yun Yeyi, both former [SenseTime](https://aiwiki.ai/wiki/sensetime) staff, as its founders.[4][7][2] MiniMax went public on the Hong Kong Stock Exchange on January 9, 2026, one day after rival [Zhipu AI](https://aiwiki.ai/wiki/zhipu_ai); its shares closed 109% above the offer price on the first day, valuing the company at about US$13.7 billion.[2][3] It is regularly grouped with Zhipu AI, [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai), [Baichuan Intelligence](https://aiwiki.ai/wiki/baichuan), [StepFun](https://aiwiki.ai/wiki/stepfun) and [01.AI](https://aiwiki.ai/wiki/01_ai) as one of China's "Six Little Tigers" of generative AI.[8] The name MiniMax is the technical term for the minimax algorithm of game theory; Yan has described the name as self-contradictory.[10]

As of December 31, 2025, MiniMax said it had cumulatively served more than 236 million users across over 200 countries and regions, and 214,000 enterprise customers and developers from more than 100 countries and regions.[1] At its August 31, 2024 partner day, the company said its models were handling an average of 3 billion AI interactions and processing more than 3 trillion text tokens per day (the English version of the announcement gives 300 trillion, a mistranslation of the Chinese 3万亿).[60][71] In the first half of 2026 its revenue rose 283.1% year on year to US$116.6 million, with the Open Platform (API and developer subscriptions) overtaking consumer apps as the largest revenue source.[22]

## What does MiniMax do?

MiniMax positions itself as a global AI [foundation model](https://aiwiki.ai/wiki/foundation_models) company that pairs proprietary research with mass-market consumer apps. Its work spans four pillars: [large language models](https://aiwiki.ai/wiki/large_language_model) (the abab, MiniMax-01, and MiniMax-M series), [AI video generation](https://aiwiki.ai/wiki/ai_video_generation) ([Hailuo AI](https://aiwiki.ai/wiki/hailuo) and the H3 model), companion and character chat (Talkie and Xingye), and speech and music generation.[1][4] Its products and API are sold through an enterprise-facing Open Platform and consumer apps.[1] On the March 2, 2026 earnings call for its first annual results as a listed company, Yan Junjie said MiniMax would evolve from a large-model company into a platform company for the AI era, and defined the value of an AI platform as "intelligence density" multiplied by token throughput.[1][23] More than 70% of MiniMax's 2025 revenue came from international markets.[1]

## History

### Founding and Early Development

MiniMax Group Inc. was incorporated in the Cayman Islands on June 30, 2021, when it was 98.5% owned by Yan Junjie. Its first Chinese operating subsidiaries, Shanghai Xiyu Jizhi Technology and Beijing Xiyu Jizhi Technology, were set up in November 2021, and the company says it began operating, with R&D staff and expenses, from the start of January 2022.[4] Chinese media, including The Paper, have dated the founding to December 2021.[68] Several early leaders came from SenseTime:

- **Yan Junjie (闫俊杰)**, founder, chairman, CEO and CTO. According to the prospectus, he spent more than six years at SenseTime, where his positions included vice president and vice-head of its research institute. He obtained a bachelor's degree in mathematics from Southeast University in June 2010, a doctorate in artificial intelligence from the Institute of Automation of the [Chinese Academy of Sciences](https://aiwiki.ai/wiki/chinese_academy_of_sciences) in July 2015, and did postdoctoral research at [Tsinghua University](https://aiwiki.ai/wiki/tsinghua_university).[4]

- **Yun Yeyi (贠烨祎)**, executive director and chief operating officer, responsible for product and commercialization. She worked at SenseTime from 2017 to January 2022, latterly as CEO executive assistant, head of strategy and director of innovative business, and joined MiniMax in March 2022. The prospectus names her and Yan as the company's founders and its two weighted-voting-rights beneficiaries.[4]

- **Yang Bin (杨斌)**, an early co-founder who led the team building MiniMax's base models, according to Chinese media. He did his PhD at the University of Toronto under [Raquel Urtasun](https://aiwiki.ai/wiki/raquel_urtasun), took part in Uber ATG's research work, and co-founded the autonomous trucking company [Waabi](https://aiwiki.ai/wiki/waabi) with Urtasun.[11] He was a director of MiniMax from December 2022 until he resigned in August 2024.[4]

- **Zhou Yucong (周彧聪)**, executive director and visual model research and engineering leader. He worked at SenseTime (April 2018 to July 2019) and Huawei (August 2019 to March 2022) before joining MiniMax in March 2022; some outlets describe him as a co-founder.[4][11]

MiniMax's first text model, abab 1, was launched in April 2022 for internal use, followed by abab 2 (June 2022), abab 3 (completed October 2022) and abab 4.[4] miHoYo, the developer of Genshin Impact, first invested in 2021, and miHoYo co-founder Liu Wei has sat on MiniMax's board as a non-executive director since April 2023.[4]

### Funding Rounds

Before its IPO, MiniMax raised about US$1.56 billion over seven rounds of preferred shares, according to its Hong Kong prospectus:[4]

| Round | Last share purchase agreement | Amount raised | Post-money valuation | Reference |
| --- | --- | --- | --- | --- |
| Series Angel | December 2, 2021 | US$31.0 million | US$200 million | [4] |
| Series Pre-A | March 29, 2022 | US$50.0 million | US$550 million | [4] |
| Series A | May 4, 2023 | US$257.0 million | US$1.157 billion | [4] |
| Series A+ | July 6, 2023 | US$50.0 million | US$1.6 billion | [4] |
| Series Pre-B | March 15, 2024 | US$654.0 million | US$2.554 billion | [4][8] |
| Series Pre-B+ | December 4, 2024 | US$123.5 million | US$3.1235 billion | [4] |
| Series Pre-B++ | August 16, 2025 | US$390.4 million | US$4.2404 billion | [4] |
| **Total (pre-IPO)** | | **about US$1.56 billion** | | |

TechNode reported that [Alibaba](https://aiwiki.ai/wiki/alibaba) led a US$600 million round in March 2024; the prospectus records the corresponding Series Pre-B round at US$654.0 million.[8][4] At the time of its listing application, Alibaba (through Alisoft China Holding) was the largest outside shareholder, with about 15.04%. Other shareholders described in the prospectus include miHoYo, IDG Capital, [Tencent](https://aiwiki.ai/wiki/tencent) (through Image Frame Investment), Hillhouse, HongShan (formerly Sequoia China), MPC VII, Shunwei-linked Astrend funds, Richard Li's Pacific Century Group, Future Capital, China Life funds, and Vitalbridge.[4]

### When did MiniMax go public? IPO and Public Listing

On July 16, 2025, Reuters reported that MiniMax had filed confidentially for a Hong Kong initial public offering, targeting a valuation of more than US$4 billion.[12] The company listed under Chapter 18C of the Listing Rules, which covers specialist technology companies, and initially as a "pre-commercial" company because it had not yet met the revenue threshold.[4] China International Capital Corporation and UBS were the joint sponsors.[4]

MiniMax priced its IPO at HK$165 per share, the top of its HK$151 to HK$165 range. After the company exercised an offer size adjustment (upsize) option, the global offering comprised 29,197,600 shares, with gross proceeds of HK$4,817.6 million and net proceeds of HK$4,596.1 million. The Hong Kong public tranche drew 417,646 valid applications and was 1,837.17 times subscribed.[5] On January 9, 2026 the underwriters also exercised the over-allotment (greenshoe) option in full, adding 4,379,640 shares at HK$165 and about HK$697.3 million in net proceeds.[6] Fourteen cornerstone investors, including the Abu Dhabi Investment Authority, Alibaba, Boyu, Mirae Asset Securities, Taikang Life and E Fund, had agreed to buy about US$350 million of shares.[4][3]

The listing entity, MiniMax Group Inc., is a Cayman Islands holding company controlled through a weighted-voting-rights structure and trades on HKEX under stock code 0100.[4] On its debut on January 9, 2026, MiniMax shares closed at HK$345, up 109% from the offer price, valuing the company at about US$13.7 billion; the shares touched HK$351.8 during the day.[3] The listing came one day after Zhipu AI's debut on the same exchange.[2] At the listing ceremony Yan said, "This is only the beginning."[3] Forbes Asia estimated that the IPO made the 36-year-old Yan worth about US$3.2 billion.[13]

On March 13, 2026, after its 2025 revenue passed the HK$250 million threshold, MiniMax announced that the exchange had approved its reclassification as a commercial company, removing the "P" marker from its stock short name.[16]

### After the listing (2026)

MiniMax shares rose sharply in the first quarter of 2026. On March 18, 2026, they jumped 22% in Hong Kong after NVIDIA CEO [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) called the open-source agent [OpenClaw](https://aiwiki.ai/wiki/openclaw) "definitely the next ChatGPT"; CNBC noted that MiniMax had recently rolled out tools built on OpenClaw.[15] According to Caixin, the stock closed at a record HK$1,238 that day and the company's market value at one point touched about HK$390 billion.[14]

On May 29, 2026, MiniMax signed a listing tutoring agreement with CITIC Securities, the first formal step toward a second listing of shares on the Shanghai Stock Exchange's STAR Market, and the CSRC Shanghai bureau accepted the tutoring filing the same day.[17][14] The shares then fell steeply. On June 1, the day MiniMax launched [MiniMax M3](https://aiwiki.ai/wiki/minimax_m3) and replaced its request-based Coding Plan subscriptions with token-metered Token Plan subscriptions, the stock fell 15.7%, its largest one-day drop since listing, closing at HK$710.[18][19] After a wave of user complaints that the change amounted to a hidden price rise, MiniMax apologised for failing to communicate the billing change in advance and announced compensation for existing subscribers.[19] MiniMax's price list now shows M3 API prices with a "permanent 50% off" discount.[61]

On July 9, 2026, the first day on which large blocks of pre-IPO shares could be sold, the stock fell 17.98% to close at HK$297.40.[21][20] The next morning, MiniMax announced a placing of 35.6 million new Class A shares at HK$268 each (gross proceeds about HK$9.54 billion) and a concurrent issue of HK$6.5 billion of zero-coupon convertible bonds due 2027 with an initial conversion price of HK$335, together about HK$16 billion; the company said it would use the money mainly for AI infrastructure and model R&D, global commercialization and development of its "Harness product".[20] Both deals were completed on July 14 and July 16, 2026.[22] On the same day as the announcement, Yan sent an all-staff letter saying he would take no salary from the company until it achieves AGI, and would give shares equal to 4% of the company's total share capital to team incentives over four years and 1% to a fund supporting open-source communities.[21]

### Financial Performance

MiniMax's revenue has grown rapidly, but the company is not yet profitable. Full-year figures for 2024 and 2025:[1]

| Metric | 2024 | 2025 | Change |
| --- | --- | --- | --- |
| Total Revenue | US$30.5 million | US$79.0 million | +158.9% |
| AI-Native Products Revenue | US$21.8 million | US$53.1 million | +143.4% |
| Open Platform & Enterprise Revenue | US$8.7 million | US$26.0 million | +197.8% |
| Gross Profit | US$3.7 million | US$20.1 million | +437.2% |
| Gross Margin | 12.2% | 25.4% | +13.2 pp |
| R&D Expenses | US$189.0 million | US$252.8 million | +33.8% |
| Adjusted Net Loss | US$244.2 million | US$250.9 million | +2.7% |
| Cash Balance (Dec 31) | US$880.6 million | US$1,050.3 million | +19.3% |

Sales and distribution expenses fell 40.3% from US$87.0 million to US$51.9 million in 2025, which MiniMax attributed to organic growth and user referrals. While the adjusted (non-IFRS) net loss was US$250.9 million, the reported IFRS net loss for 2025 was US$1,871.6 million, driven mainly by a US$1,589.9 million non-cash fair value loss on its convertible redeemable preferred shares as the company's valuation rose.[1] On the March 2026 earnings call, Yan said the company's annualized recurring revenue (ARR) had passed US$150 million in February 2026.[23]

The first half of 2026 (six months to June 30):[22]

| Metric | H1 2025 | H1 2026 | Change |
| --- | --- | --- | --- |
| Total Revenue | US$30.4 million | US$116.6 million | +283.1% |
| Open Platform & Enterprise Revenue | US$9.2 million | US$73.9 million | +703.1% |
| AI-Native Products Revenue | US$21.2 million | US$42.6 million | +100.9% |
| Gross Margin | 12.1% | 17.9% | +5.8 pp |
| R&D Expenses | US$124.3 million | US$296.9 million | +138.8% |
| Loss for the Period (IFRS) | US$402.2 million | US$358.0 million | -11.0% |
| Adjusted Net Loss | US$138.7 million | US$293.0 million | +111.2% |

MiniMax attributed the Open Platform growth to more paying users and enterprise customers, higher API call volumes and "the rapid adoption of our Token Plan"; that segment made up 63.4% of revenue, against 30.3% a year earlier. Revenue from outside mainland China was 60.8% of the total. Cash balance was US$1,322.8 million at June 30, 2026, before the July placing and bond issue.[22]

## Products and Services

### Consumer Applications

#### Glow and Its Successors

MiniMax's first consumer product was **Glow**, a chat app launched in October 2022 that let users talk to AI characters.[24] The company's prospectus says an experimental dialogue product it launched in October 2022 had about 100,000 users by the end of 2022.[4] Glow was removed from Chinese app stores in March 2023, before China had issued its rules for registering generative AI services.[24] MiniMax then launched two successors:

- **Talkie** for international markets (June 2023)[4]

- **Xingye (星野)** for the Chinese market (September 2023). PingWest reported that it went live on September 8, 2023, shortly after MiniMax's model became one of the first in China to pass the generative AI registration on August 31, 2023.[4][24]

#### Talkie

Talkie is a companion AI chatbot app that lets users create and talk to virtual characters based on fiction or real people.[25] According to Sensor Tower data cited by the South China Morning Post, Talkie was the fourth most-downloaded AI app in the United States in the first half of 2024, ahead of Google-backed [Character.ai](https://aiwiki.ai/wiki/character_ai) in 10th place, and recorded 17 million global downloads in the first eight months of 2024, against nearly 19 million for Character.ai. Sensor Tower also found that users aged 18 to 35 made up more than 70% of the user base of the top companion apps, a group that included Talkie and Character.ai.[25]

In December 2024 Talkie was removed from Apple's US App Store; the app attributed the removal to unspecified "technical reasons", while it remained available on Google Play.[26] In the fourth quarter of 2025 MiniMax released M2-her, a model built to power Talkie and Xingye.[1]

#### Hailuo AI (海螺AI)

**Hailuo AI** started as a text-based productivity assistant; Yan later said it first launched about two years before January 2025 but failed to take off at first.[10] In April 2024 MiniMax relaunched it on the abab 6.5 model, with voice calls, text chat and image recognition.[27] After MiniMax added its first video model on August 31, 2024, the product drew a surge of overseas users, and the prospectus now describes Hailuo AI as MiniMax's visual generation platform.[27][4] By the end of 2025 MiniMax said its video models had helped creators generate more than 600 million videos.[1] Hailuo competes with [Runway](https://aiwiki.ai/wiki/runway_ml), [Kling](https://aiwiki.ai/wiki/kling) from [Kuaishou](https://aiwiki.ai/wiki/kuaishou) and Google's [Veo](https://aiwiki.ai/wiki/veo), and earlier competed with [OpenAI](https://aiwiki.ai/wiki/openai)'s [Sora](https://aiwiki.ai/wiki/sora), which OpenAI is shutting down in 2026 (the app closed in April; the API was scheduled to end on September 24).[72] Key releases include:

| Model | Type | Release | Key Capabilities |
| --- | --- | --- | --- |
| Video-01 (Hailuo-01) | Text-to-video and image-to-video | August 31, 2024 | 720p at 25 fps, up to 6 seconds, cinematic camera movement[28] |
| S2V-01 | Subject-to-video | January 2025 | Character-consistent video from a single reference image[29] |
| I2V-01-Live | Image-to-video (animation) | January 2025 | Smoother animation for 2D illustrations[30] |
| T2V-01-Director | Text-to-video (director mode) | January 28, 2025 | Camera control through natural language or commands, less random movement[31][32] |
| I2V-01-Director | Image-to-video (director mode) | Early 2025 | Director-mode camera control for image-to-video[32] |
| Hailuo 02 | Next-gen video | June 18, 2025 | Native 1080p, Noise-aware Compute Redistribution (NCR) architecture, 3x parameters and 4x training data versus its predecessor[33] |
| Hailuo 2.3 | Enhanced video | October 28, 2025 | Better physical actions, stylization and micro-expressions; Hailuo 2.3 Fast cuts batch creation costs by up to 50%[34] |
| [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3) | Omni-modal video | July 31, 2026 (weights August 3, 2026) | Up to 15 seconds at 2K with native stereo audio from mixed text, image, video and audio context[35][36] |

MiniMax says the NCR framework behind Hailuo 02 boosts training and inference efficiency by 2.5 times at comparable parameter scale. Hailuo 02 shipped in three variants: 768p for 6 seconds, 768p for 10 seconds, and 1080p for 6 seconds.[33]

In summer 2025 Hailuo introduced a Video Agent, which evolved into a "Media Agent" alongside the Hailuo 2.3 launch in October 2025. The Media Agent automatically picks the multimodal models needed for a request and offers "one-click video generation" without manual editing.[34]

In July 2026 the video line moved to [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3), an omni-modal model that generates up to 15-second 2K clips with native stereo audio; MiniMax released its weights on August 3, 2026, and its platform documentation now lists the Hailuo models as legacy models.[35][36][62]

#### MiniMax Agent and MiniMax Code

MiniMax launched MiniMax Agent, a general-purpose agent for long-horizon tasks, on June 19, 2025, saying it had already been used internally for nearly 60 days.[37] In January 2026 the company launched MiniMax Agent as an "AI-native workspace".[1] The desktop app was renamed [MiniMax Code](https://aiwiki.ai/wiki/minimax_code) on May 29, 2026, and MiniMax Code 2.0, rebuilt on the open-source Pi agent framework, shipped in the desktop app on July 13, 2026 and was announced globally on August 7, 2026.[38][39]

### AI Models

MiniMax has developed a succession of models across text, vision, speech, and music.

#### abab Series

The abab series was MiniMax's first generation of [large language models](https://aiwiki.ai/wiki/large_language_model):

| Model | Type | Release | Key Features |
| --- | --- | --- | --- |
| abab 1 | [LLM](https://aiwiki.ai/wiki/large_language_model) | April 2022 | First text model, for MiniMax's own products[4] |
| abab 2 to abab 4 | LLM | June to late 2022 | Successive internal models; abab 3 completed October 2022[4] |
| abab 5.5 | LLM | May 2023 | Launched as MiniMax signed its first API customer[4] |
| abab 6 | [Mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts) (MoE) | January 2024 | MoE text model; by the company's own evaluations below [GPT-4](https://aiwiki.ai/wiki/gpt-4) but ahead of GPT-3.5 on some abilities[4][40] |
| abab 6.5 / 6.5s | Mixture of experts | April 17, 2024 | abab 6.5 has about a trillion parameters and a 200K-token context; abab 6.5s processes nearly 30,000 words per second[41] |

#### MiniMax-01 Series (Lightning Attention)

The MiniMax-01 series, released and open-sourced on January 15, 2025, introduced a hybrid architecture combining [Lightning Attention](https://aiwiki.ai/wiki/lightning_attention), softmax [attention](https://aiwiki.ai/wiki/attention), and [Mixture-of-Experts](https://aiwiki.ai/wiki/mixture_of_experts) (MoE), with weights and code published on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) and GitHub.[42][43]

**Lightning Attention** is MiniMax's core technical innovation. Traditional attention mechanisms scale quadratically with sequence length, making very long context windows computationally expensive. Lightning Attention is a linear attention implementation that splits the calculation into intra-block computations (using left-product attention) and inter-block computations (using right-product attention). The MiniMax-01 paper argues that its capacity is O(d^2/h), larger than softmax attention's O(d), where d is the model dimension and h the number of heads.[42]

To train these models, MiniMax developed expert tensor parallel (ETP) and an EP-ETP overlap strategy for MoE, varlen ring attention, and an improved version of Linear Attention Sequence Parallelism (LASP).[42]

| Model | Type | Parameters | Context Window | Key Features |
| --- | --- | --- | --- | --- |
| MiniMax-Text-01 | Text-only LLM | 456 billion total (45.9B active per token) | 1 million (training), up to 4 million (inference) | 32 MoE experts, hybrid Lightning/softmax attention, open-weight[42] |
| MiniMax-VL-01 | Multimodal LLM | Built on MiniMax-Text-01 | | Continued training on 512 billion vision-language tokens[42] |

The MiniMax-01 paper reports that the models match state-of-the-art models such as [GPT-4o](https://aiwiki.ai/wiki/gpt_4o) and [Claude 3.5 Sonnet](https://aiwiki.ai/wiki/claude_3_5_sonnet) on standard and in-house benchmarks while offering a context window 20 to 32 times longer.[42]

#### MiniMax-M1 (Hybrid-Attention Reasoning)

[MiniMax-M1](https://aiwiki.ai/wiki/minimax_m1), released on June 16, 2025, is described by MiniMax as the world's first open-weight, large-scale hybrid-attention reasoning model. Built on MiniMax-Text-01 (456 billion total parameters, 45.9 billion active), it combines Lightning Attention with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning)-based reasoning and natively supports a 1-million-token context window, eight times DeepSeek R1's.[44] It was released in 40K and 80K thinking-budget versions under the Apache 2.0 license.[44][56]

MiniMax says the full reinforcement learning phase for M1 ran on 512 NVIDIA H800 GPUs for three weeks, at a rental cost of US$534,700, helped by a new RL algorithm called CISPO. The company also says that when generating 80,000 tokens of reasoning, M1 needs only about 30% of the computing power of rival [DeepSeek](https://aiwiki.ai/wiki/deepseek)'s R1.[44][45]

#### MiniMax-M2 Series

[MiniMax-M2](https://aiwiki.ai/wiki/minimax_m2), open-sourced on October 27, 2025, is a compact MoE model built for coding and agents, with 230 billion total parameters and 10 billion active parameters per token (229.9B and 9.8B in its technical report). MiniMax priced its API at US$0.30 per million input tokens and US$1.20 per million output tokens, which it said was 8% of the price of Claude Sonnet 4.5.[46][49] The weights are released under an MIT license whose one added condition requires products with more than 100 million monthly active users or US$30 million in annual recurring revenue to display "MiniMax M2" prominently.[48] Artificial Analysis said the model could be served on as few as four [NVIDIA](https://aiwiki.ai/wiki/nvidia) H100 GPUs at FP8 precision.[47]

| Model | Release | Key Features |
| --- | --- | --- |
| MiniMax-M2 | October 27, 2025 | 230B total / 10B active parameters; coding, tool use and deep search[49][1] |
| MiniMax-M2.1 | December 23, 2025 | Stronger multi-language programming and complex real-world tasks[50] |
| MiniMax-M2-her | Fourth quarter of 2025 | Dialogue model behind Talkie and Xingye[1] |
| MiniMax-M2.5 | February 12, 2026 | Vendor-reported 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp; completes SWE-Bench Verified 37% faster than M2.1[51] |
| [MiniMax-M2.7](https://aiwiki.ai/wiki/minimax_m2_7) | March 18, 2026 | "Self-evolving" model; vendor-reported 56.22% on SWE-Pro and 57.0% on Terminal Bench 2[52] |

According to Artificial Analysis data reported by VentureBeat, M2 scored 61 on the firm's intelligence index at launch, the highest of any open-weight model and close behind [GPT-5](https://aiwiki.ai/wiki/gpt-5) (high) and Grok 4. VentureBeat added that MiniMax's own agentic benchmark results put it at or near proprietary systems such as GPT-5 (thinking) and Claude Sonnet 4.5.[47] MiniMax said M2 became the first Chinese model on OpenRouter to exceed 50 billion tokens of daily consumption.[1]

MiniMax-M2.7, released on March 18, 2026, is described as "self-evolving": MiniMax said it used an internal version of the model to build and run agent harnesses for its reinforcement learning team, where M2.7 handled 30% to 50% of the workflow, and that in one test M2.7 autonomously optimized a coding scaffold over more than 100 iterative rounds.[52] When MiniMax published M2.7's weights on Hugging Face in April 2026, it used a license that allows non-commercial use on MIT-style terms but prohibits commercial use without separate prior written authorization, a shift from the M2 license that drew complaints from some developers on the model's Hugging Face page.[53][54][69]

#### MiniMax M3

[MiniMax M3](https://aiwiki.ai/wiki/minimax_m3), released on June 1, 2026, is a natively multimodal model with a 1-million-token context, about 428 billion total parameters and about 23 billion activated parameters. It uses MiniMax Sparse Attention (MSA), which the company says gives more than 9 times faster prefill and more than 15 times faster decoding than its previous-generation model at 1 million tokens of context; the MSA paper reports 14.2x prefill and 7.6x decoding speedups on a 109-billion-parameter test model.[55][56][57] MiniMax reported 59.0% on [SWE-Bench Pro](https://aiwiki.ai/wiki/swe_bench_pro) and 66.0% on [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1.[55] The weights were published on Hugging Face in June 2026 under a MiniMax Community License that requires a "Built with MiniMax M3" notice for commercial use and separate authorization for products with more than US$20 million in yearly revenue.[56]

#### Speech Models

| Model | Release | Key Features |
| --- | --- | --- |
| Speech-01 | November 2023 | MiniMax's first speech model[4] |
| Speech-02 | April 2025 | Autoregressive Transformer text-to-speech with a learnable speaker encoder; 32 languages[4][58] |
| Speech 2.6 | October 2025 | Optimized for voice agents; more than 40 languages[1] |

MiniMax says Speech-02, listed as Speech-02-HD, ranked first on both the [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) Speech Arena and the Hugging Face TTS Arena.[58] By the end of 2025 the company said its speech models had generated more than 200 million hours of speech.[1] Its current API lineup is Speech 2.8 in HD and Turbo variants.[62]

#### Music Models

**Music-01**, introduced on August 31, 2024, generates accompaniment and vocals together. Users can upload reference music for the model to learn rhythm and style, then input lyrics to receive a new AI-generated piece. At launch it handled classical, pop, rock, electronic and a dozen more styles, with a maximum output length of 60 seconds.[59] MiniMax released Music 2.0 in October 2025, and says Music 2.0 and 2.5 can compose songs of up to five minutes.[4][1] From August 20, 2026, MiniMax closed its paid music generation APIs to new users (existing paying users kept access) and discontinued its free music APIs, directing users instead to its MiniMax Audio app and to the open-weight MiniMax Music 3 model it had published on Hugging Face.[62][61]

### Comprehensive Model Timeline

| Model Name | Type | Release Date | Key Features |
| --- | --- | --- | --- |
| abab 1 | LLM | April 2022 | First text model |
| abab 5.5 | LLM | May 2023 | First API customer |
| Speech-01 | Text-to-speech | November 2023 | First speech model |
| abab 6 | MoE LLM | January 2024 | MoE architecture |
| abab 6.5 | MoE LLM | April 2024 | About a trillion parameters, 200K context |
| Music-01 | Music generation | August 2024 | Simultaneous vocals and accompaniment |
| Video-01 | Video generation | August 2024 | 720p, 6 seconds |
| MiniMax-Text-01 | MoE LLM | January 2025 | 456B params, 4M token inference context, Lightning Attention |
| MiniMax-VL-01 | Multimodal LLM | January 2025 | Vision + text understanding |
| Speech-02 | Text-to-speech | April 2025 | 32 languages |
| MiniMax-M1 | Reasoning LLM | June 2025 | Hybrid-attention reasoning, 1M context |
| Hailuo 02 | Video generation | June 2025 | Native 1080p, NCR architecture |
| MiniMax-M2 | Open-weight LLM | October 2025 | 230B/10B active |
| Hailuo 2.3 | Video generation | October 2025 | Improved physics, stylization, micro-expressions |
| MiniMax-M2.1 | Open-weight LLM | December 2025 | Multi-language programming |
| MiniMax-M2.5 | Open-weight LLM | February 2026 | Productivity and agentic tasks |
| MiniMax-M2.7 | Open-weight LLM (non-commercial license) | March 2026 | Self-evolving, automated RL research |
| MiniMax M3 | Open-weight multimodal LLM | June 2026 | MSA sparse attention, 1M context |
| MiniMax H3 | Open-weight video model | July/August 2026 | 2K, 15 seconds, native stereo audio |

### API and Enterprise Services

MiniMax provides API access through its Open Platform (platform.minimax.io), which served 214,000 enterprise customers and developers as of December 2025.[1] The platform offers text, vision, video and speech models (its music APIs closed to new users on August 20, 2026), sold pay-as-you-go or through Token Plan subscriptions.[61][62] List prices for its language models as of September 2026:[61]

| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
| --- | --- | --- |
| MiniMax-M3 (up to 512K input tokens, after permanent 50% discount) | $0.30 | $1.20 |
| MiniMax-M3 (above 512K input tokens, after permanent 50% discount) | $0.60 | $2.40 |
| MiniMax-M2.7 | $0.30 | $1.20 |
| MiniMax-M2.5 | $0.30 | $1.20 |
| MiniMax-M2 | $0.30 | $1.20 |

Demand for the M-series accelerated after the IPO: MiniMax reported that average daily token consumption of its M2-series text models in February 2026 was more than six times the December 2025 level, and that token consumption from its Coding Plan grew more than tenfold over the same period.[1]

## Corporate Structure

The publicly listed entity, MiniMax Group Inc., is a Cayman Islands holding company controlled through weighted voting rights: each Class B share carries ten votes, and Yan Junjie and Yun Yeyi are the weighted-voting-rights beneficiaries.[4] Its major operating subsidiaries, all wholly owned, include Shanghai Xiyu Jizhi Technology Co., Ltd. (上海稀宇极智科技有限公司, founded November 3, 2021) and Beijing Xiyu Jizhi Technology (November 18, 2021), which do foundation model research; SUBSUP Pte. Ltd. in Singapore, which operates AI-native products; Shanghai Xiyu Technology Co., Ltd. (上海稀宇科技有限公司, January 2023); Nanonoble Pte. Ltd. in Singapore (March 2024); and MiniMax HongKong (April 2025).[4] Shanghai Xiyu Technology was previously controlled through contractual arrangements because of limits on foreign ownership of internet services, and was restructured before the listing.[4]

## Market Position

### China's "Six Little Tigers" of AI

MiniMax is one of China's "Six Little Tigers" (六小虎), a label for the leading Chinese AI startups focused on large language models that emerged around the launch of [ChatGPT](https://aiwiki.ai/wiki/chatgpt). The six companies are:[8]

| Company | Focus | Notable Products |
| --- | --- | --- |
| **MiniMax** | Multimodal AI, consumer apps | Talkie, Hailuo AI, abab/M-series models |
| [Zhipu AI](https://aiwiki.ai/wiki/zhipu_ai) | General LLMs, enterprise AI | GLM series, ChatGLM |
| [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai) | Long-context AI | Kimi chatbot |
| [Baichuan Intelligence](https://aiwiki.ai/wiki/baichuan) | General LLMs | Baichuan series |
| [01.AI](https://aiwiki.ai/wiki/01_ai) | LLMs | Yi series |
| [StepFun](https://aiwiki.ai/wiki/stepfun) (阶跃星辰) | Multimodal AI | Step series |

In April 2024, the South China Morning Post described four "AI tigers": Baichuan, Zhipu AI, Moonshot AI and MiniMax.[9] MiniMax and Zhipu AI became the first two of the group to list, on consecutive days in January 2026.[3]

### Competitive Landscape

MiniMax competes across multiple AI segments:

- **Character AI/Companion chatbots:** Talkie competes with [Character.ai](https://aiwiki.ai/wiki/character_ai) and other [AI companion](https://aiwiki.ai/wiki/ai_companions) platforms. In the first half of 2024, Talkie ranked above Character.ai in U.S. AI app downloads.[25]
- **Video generation:** Hailuo AI and H3 compete with [Runway](https://aiwiki.ai/wiki/runway_ml), [Kling](https://aiwiki.ai/wiki/kling) (Kuaishou) and Google [Veo](https://aiwiki.ai/wiki/veo); OpenAI's [Sora](https://aiwiki.ai/wiki/sora) was a rival until OpenAI began shutting it down in 2026.[72]
- **Large language models:** The MiniMax-M series competes with [GPT-5](https://aiwiki.ai/wiki/gpt-5) ([OpenAI](https://aiwiki.ai/wiki/openai)), [Claude](https://aiwiki.ai/wiki/claude) ([Anthropic](https://aiwiki.ai/wiki/anthropic)), [Gemini](https://aiwiki.ai/wiki/gemini) ([Google](https://aiwiki.ai/wiki/google)), [DeepSeek](https://aiwiki.ai/wiki/deepseek), [Qwen](https://aiwiki.ai/wiki/qwen), and [Llama](https://aiwiki.ai/wiki/llama) ([Meta](https://aiwiki.ai/wiki/meta)).
- **Text-to-speech:** MiniMax's speech models compete with [ElevenLabs](https://aiwiki.ai/wiki/eleven_labs) and OpenAI's speech models.
- **Music generation:** MiniMax's music models compete with [Suno](https://aiwiki.ai/wiki/suno) and [Udio](https://aiwiki.ai/wiki/udio).

### Industry Recognition

On July 16, 2025, speaking at the opening ceremony of a supply chain expo in Beijing, [NVIDIA](https://aiwiki.ai/wiki/nvidia) CEO Jensen Huang said: "Models like DeepSeek, Alibaba, Tencent, MiniMax, and Baidu Ernie bot are world class, developed here and shared openly [and] have spurred AI developments worldwide."[67]

### Industry Outlook

During the 2024 World Artificial Intelligence Conference, Yan Junjie told reporters that he expected only five large-model companies to remain worldwide, and that in the AI era the split of the market between big tech companies and startups could reach an extreme ratio of 9:1.[68] In the March 2026 results release he predicted that coding models would reach "L4 to L5 levels of intelligence" during 2026, shifting AI from a tool to a colleague-level partner.[1]

## Technical Infrastructure

### Lightning Attention Architecture

MiniMax's best-known technical contribution is Lightning Attention, a linear-complexity attention mechanism for very long sequences. Published with the MiniMax-01 release (arXiv: 2501.08313), it splits the attention computation into intra-block and inter-block operations:[42]

- **Intra-block computation** uses left-product attention for tokens within the same block.
- **Inter-block computation** uses the right-product kernel trick for tokens across different blocks, keeping the computational cost linear with respect to sequence length.

In MiniMax-Text-01, a transformer block with softmax attention follows every seven blocks of linear (lightning) attention, for 80 layers in total. This lets the model train on contexts of 1 million tokens and extrapolate to 4 million tokens during inference.[42] With M3 in 2026, MiniMax moved to a different design, MiniMax Sparse Attention, a blockwise sparse attention built on grouped-query attention that selects a top-k subset of key-value blocks for each query group.[57]

### Is MiniMax open source?

MiniMax has been one of the more prominent open-weight developers among the Six Little Tigers, though its licensing has tightened over time. In January 2025 it released the MiniMax-01 model family with weights and code.[43] In a LatePost interview published that month, Yan said that if he could choose again, MiniMax should have open-sourced from the first day, because open source accelerates technical progress.[10] MiniMax-M1 used the Apache 2.0 license, and the M2 series used an MIT-based license with an attribution clause for very large deployments.[44][48] MiniMax-M2.7 moved to a license that bars commercial use without written authorization, and M3 and H3 use custom community licenses with commercial conditions.[53][56][70] In July 2026 Yan pledged shares equal to 1% of the company's share capital to a fund supporting open-source communities.[21]

### Parallelism and Training Innovations

MiniMax has developed several training optimization techniques for distributed computing at scale:[42][44]

- **LASP+ (improved Linear Attention Sequence Parallelism):** A sequence-parallelism strategy for linear attention layers.
- **Varlen Ring Attention:** A variable-length ring attention mechanism that reduces redundant computation when packing sequences of different lengths.
- **ETP (Expert Tensor Parallel):** A parallelism method for Mixture-of-Experts layers, combined with an EP-ETP overlap strategy.
- **Compute-Communication Overlap:** Methods to overlap computation and communication during distributed training.
- **CISPO:** An RL algorithm introduced with M1 that clips importance sampling weights rather than token updates.

## Controversies

### Hollywood Copyright Lawsuit

On September 16, 2025, Disney, Universal and Warner Bros. Discovery filed a copyright infringement lawsuit in the U.S. District Court for the Central District of California over Hailuo AI.[4][63] The defendants are Shanghai Xiyu Jizhi Technology and the Singapore-based Nanonoble, which the studios said operate under the MiniMax name.[63][64] The complaint called MiniMax's infringement "willful and brazen" and said MiniMax markets Hailuo AI as a "Hollywood studio in your pocket," which it called "an audacious self-anointed nickname given that MiniMax built its business from intellectual property stolen from Hollywood studios like Plaintiffs."[63] In its prospectus MiniMax said its directors believed the claims were without merit.[4] In May 2026 the court denied the defendants' motion to dismiss, holding that they were subject to U.S. personal jurisdiction (citing their application to register the MINIMAX trademark with the U.S. Patent and Trademark Office) and that the studios had plausibly alleged infringement both in training and in outputs.[64] MiniMax later cited "ongoing copyright-related legal proceedings specifically concerning generative video AI" as one reason it excluded the United States from the license territory of its open-weight H3 model.[70]

In December 2024, Broadcast reported that Hailuo AI generated videos showing what appeared to be the logos of British broadcasters ITV, Channel 4 and Channel 5, and that ITV was investigating.[65]

### Other Legal Issues

In February 2025, Jiemian reported that iQiyi, the Chinese video streaming service, had sued MiniMax in the Xuhui District People's Court in Shanghai, alleging that MiniMax used iQiyi's copyrighted audiovisual works to train its models without authorization; iQiyi sought an end to the infringement and damages of about 100,000 yuan.[66]

### M3 Billing Change

The switch from request-based Coding Plan subscriptions to token-metered Token Plan subscriptions at the June 1, 2026 launch of M3 drew a wave of complaints from users who said they had not been told in advance and that the same tasks now used up their monthly quotas far faster. MiniMax apologised, saying it had not communicated the changes adequately, and offered measures including continued unlimited weekly quotas for some early subscribers, a permanent 50% increase in weekly M3 quotas for others, doubled usage limits for the first week, longer validity for compensation credits, and faster work on a self-service refund channel.[19]

## See also

- [Reactor](https://aiwiki.ai/wiki/reactor_ai)
- [PixVerse](https://aiwiki.ai/wiki/pixverse)
- [Artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence)
- [Large language model](https://aiwiki.ai/wiki/large_language_model)
- [SenseTime](https://aiwiki.ai/wiki/sensetime)
- [DeepSeek](https://aiwiki.ai/wiki/deepseek)
- [Generative artificial intelligence](https://aiwiki.ai/wiki/generative_ai)
- [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai)
- [Zhipu AI](https://aiwiki.ai/wiki/zhipu_ai)
- [01.AI](https://aiwiki.ai/wiki/01_ai)
- [Character.ai](https://aiwiki.ai/wiki/character_ai)
- [Mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts)
- [Attention mechanism](https://aiwiki.ai/wiki/attention)
- [Yan Junjie](https://aiwiki.ai/wiki/yan_junjie)

## References

[1] "MiniMax Global Announces Full Year 2025 Financial Results." MiniMax News, March 2, 2026. https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results

[2] "MiniMax doubles in Hong Kong debut, marking yet another Chinese AI listing." CNBC, January 9, 2026. https://www.cnbc.com/2026/01/09/minimax-hong-kong-ipo-ai-tigers-zhipu.html

[3] "MiniMax, China's second 'AI tiger' to go public, doubles in value in Hong Kong debut." Reuters (via Yahoo Finance), January 9, 2026. https://finance.yahoo.com/news/china-ai-firm-minimax-set-012942121.html

[4] MiniMax Group Inc. "Global Offering" (prospectus). Hong Kong Exchanges and Clearing, December 31, 2025. https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1231/2025123100025.pdf

[5] MiniMax Group Inc. "Announcement of Final Offer Price and Allotment Results." HKEX, January 8, 2026. https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0108/2026010801342.pdf

[6] MiniMax Group Inc. "Full Exercise of the Over-allotment Option." HKEX, January 11, 2026. https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0111/2026011100015.pdf

[7] "About." MiniMax (company website). https://www.minimax.io/about

[8] "Meet China's top six AI unicorns: who are leading the wave of AI in China." TechNode, January 9, 2025. https://technode.com/2025/01/09/meet-chinas-top-six-ai-unicorns-who-are-leading-the-wave-of-ai-in-china/

[9] "China's 4 new 'AI tigers' - Baichuan, Zhipu AI, Moonshot AI and MiniMax - emerge as investor favourites." South China Morning Post, April 19, 2024. https://www.scmp.com/tech/big-tech/article/3259499/chinas-four-new-ai-tigers-baichuan-zhipu-ai-moonshot-ai-and-minimax-emerge-investor-favourites

[10] "MiniMax创始人闫俊杰：创业没有天选之子" (MiniMax founder Yan Junjie: there are no chosen ones in entrepreneurship). LatePost (晚点LatePost) via Tencent News, January 17, 2025. https://news.qq.com/rain/a/20250117A04P6600

[11] "获阿里领投 6 亿美元的 AI 公司 MiniMax 联合创始人杨斌被曝离职，去向暂时未知" (MiniMax co-founder Yang Bin reported to have left). IT之家 via Sina, July 6, 2024. https://tech.sina.cn/2024-07-06/detail-inccfaic3009613.d.html

[12] "Chinese AI firm MiniMax targets $4 billion-plus valuation in Hong Kong IPO, sources say." Reuters (via Yahoo Finance), July 16, 2025. https://finance.yahoo.com/news/chinese-ai-firm-minimax-files-083423828.html

[13] Forbes Asia (@ForbesAsia). Post on Yan Junjie's net worth after the MiniMax IPO. X, January 12, 2026. https://x.com/ForbesAsia/status/2010834299449168286

[14] "MiniMax启动A股科创板IPO 新模型大幅涨价" (MiniMax starts STAR Market IPO; new model sharply more expensive). Caixin (财新网), June 1, 2026. https://m.caixin.com/m/2026-06-01/102449721.html

[15] "China's 'AI tigers' see shares surge after Nvidia CEO touts OpenClaw as 'next ChatGPT'." CNBC, March 18, 2026. https://www.cnbc.com/2026/03/18/china-ai-zhipu-minimax-after-nvidia-jensen-huang-openclaw-comments.html

[16] MiniMax Group Inc. "(1) The Stock Exchange's Approval of the Company as a Commercial Company; (2) Removal of Stock Marker." HKEX, March 13, 2026. https://www.hkexnews.hk/listedco/listconews/sehk/2026/0313/2026031301613.pdf

[17] "关于MiniMax Group Inc.首次公开发行股票并在上市辅导备案报告" (Tutoring filing report for MiniMax Group Inc.'s initial public offering). CITIC Securities, June 1, 2026. https://www.cs.ecitic.com/newsite/tzgg/ipoqyfdgg/202606/t20260601_1212646.html

[18] "MiniMax港股收跌15.7% 创上市以来最大单日跌幅" (MiniMax closes down 15.7% in Hong Kong, its biggest one-day drop since listing). 观点网 via Sina Finance, June 1, 2026. https://finance.sina.com.cn/stock/estate/integration/2026-06-01/doc-inhzwyqq3928508.shtml

[19] "更改用户套餐引发大量投诉，MiniMax致歉" (Plan changes trigger complaints; MiniMax apologises). Sina Tech, June 5, 2026. https://finance.sina.com.cn/tech/roll/2026-06-05/doc-iniaiyyr1649729.shtml

[20] MiniMax Group Inc. "Placing of New Class A Shares under General Mandate and Proposed Issue of HK$6,500 Million Zero Coupon Guaranteed Convertible Bonds due 2027." HKEX, July 10, 2026. https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0710/2026071000027.pdf

[21] "MiniMaxCEO闫俊杰发全员信：不再领取薪酬，个人将拿出4%股份用于团队激励，拿出1%股份支持开源发展" (MiniMax CEO Yan Junjie's all-staff letter: no salary, 4% of shares for team incentives, 1% for open source). Securities Times (证券时报), from National Business Daily, July 10, 2026. https://www.stcn.com/article/detail/4010839.html

[22] MiniMax Group Inc. "Interim Results Announcement for the Six Months Ended June 30, 2026." August 26, 2026. https://file.cdn.minimax.io/public/586c7995-cab1-41b5-9178-c06685b32c88.pdf

[23] "MiniMax开年强劲增长 最新ARR已超1.5亿美元" (MiniMax grows strongly at start of year; ARR exceeds US$150 million). Securities Daily (证券日报网), March 4, 2026. http://www.zqrb.cn/gscy/qiyexinxi/2026-03-04/A1772627093741.html

[24] "制造星野：穿越平行时空和次元壁，揭秘 MiniMax 的另类 AI 世界" (Making Xingye: inside MiniMax's AI world). PingWest (品玩), May 7, 2024. https://www.pingwest.com/a/294725

[25] "Chinese AI unicorn MiniMax scores big in US with Talkie chatbot entertainment app." South China Morning Post (via Yahoo Finance), October 31, 2024. https://finance.yahoo.com/news/chinese-ai-unicorn-minimax-scores-093000101.html

[26] "Chinese-owned Character.ai rival vanishes from US App Store." South China Morning Post, December 20, 2024. https://www.scmp.com/tech/tech-trends/article/3291715/chinese-owned-characterai-rival-vanishes-us-app-store

[27] "国产AI出海跑出黑马，MiniMax借海螺AI爆火海外" (MiniMax's Hailuo AI takes off overseas). 21st Century Business Herald (21世纪经济报道), October 23, 2024. https://www.21jingji.com/article/20241023/herald/244f291c7e2cd0ee7bec2ade12ba2403.html

[28] "MiniMax officially releases the Video-01 video generation model." MiniMax News, August 31, 2024. https://www.minimax.io/news/video-01

[29] "S2V-01: Subject Reference of Hailuo." MiniMax News, January 2025. https://www.minimax.io/news/s2v-01-release

[30] "I2V-01-live: smooth, vivid, and perfect for 2D illustrations." MiniMax News, January 2025. https://www.minimax.io/news/i2v-01-live-release

[31] Hailuo AI (@Hailuo_AI). "Introducing Hailuo T2V-01-Director Model: Control Your Camera Like a Pro!" X, January 28, 2025. https://x.com/Hailuo_AI/status/1884176446702428568

[32] "Hailuo AI Advances Cinematic Storytelling with T2V-01-Director and I2V-01-Director." MiniMax News. https://www.minimax.io/news/01-director

[33] "MiniMax Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency." MiniMax News, June 18, 2025. https://www.minimax.io/news/minimax-hailuo-02

[34] "MiniMax Hailuo 2.3: A New Level of Complex Video Performance & Media Agent." MiniMax News, October 28, 2025. https://www.minimax.io/news/minimax-hailuo-23

[35] "MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities." MiniMax blog, July 31, 2026. https://www.minimax.io/blog/minimax-h3

[36] "Open General Intelligence: MiniMax H3 Is Now Open Source." MiniMax News, August 3, 2026. https://www.minimax.io/news/minimax-h3-open-source

[37] "MiniMax Agent: Code is Cheap, Show Me the Requirement." MiniMax News, June 19, 2025. https://www.minimax.io/news/minimax-agent

[38] "Changelog." MiniMax Agent Docs (MiniMax Code release notes). https://agent.minimax.io/docs/changelog

[39] MiniMax_Agent (@MiniMaxAgent). MiniMax Code 2.0 launch announcement. X, August 7, 2026. https://x.com/MiniMaxAgent/status/2085679716976243045

[40] "MoE架构，全量上线，MiniMax的abab6跨过玫瑰三段论了吗？" (MoE architecture, full launch: has MiniMax's abab6 made it?). PingWest (品玩), January 19, 2024. https://www.pingwest.com/a/291942

[41] "The General Large Language Model abab6.5 Series." MiniMax News, April 17, 2024. https://www.minimax.io/news/abab65-series

[42] MiniMax et al. "MiniMax-01: Scaling Foundation Models with Lightning Attention." arXiv:2501.08313, January 2025. https://arxiv.org/abs/2501.08313

[43] "MiniMax-01 is Now Open-Source: Scaling Lightning Attention for the AI Agent Era." MiniMax News, January 15, 2025. https://www.minimax.io/news/minimax-01-series-2

[44] MiniMax et al. "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention." arXiv:2506.13585, June 2025. https://arxiv.org/abs/2506.13585

[45] "MiniMax-M1, the World's First Open-Source, Large-Scale, Hybrid-Attention Reasoning Model." MiniMax News, June 16, 2025. https://www.minimax.io/news/minimaxm1

[46] "MiniMax M2 & Agent: Ingenious in Simplicity." MiniMax News, October 27, 2025. https://www.minimax.io/news/minimax-m2

[47] "MiniMax-M2 is the new king of open source LLMs (especially for agentic tool calling)." VentureBeat, October 27, 2025. https://venturebeat.com/ai/minimax-m2-is-the-new-king-of-open-source-llms-especially-for-agentic-tool

[48] "LICENSE." MiniMax-AI/MiniMax-M2, GitHub. https://github.com/MiniMax-AI/MiniMax-M2/blob/main/LICENSE

[49] "The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence." arXiv:2605.26494, May 2026. https://arxiv.org/abs/2605.26494

[50] "MiniMax M2.1: Significantly Enhanced Multi-Language Programming, Built for Real-World Complex Tasks." MiniMax News, December 23, 2025. https://www.minimax.io/news/minimax-m21

[51] "MiniMax M2.5: Built for Real-World Productivity." MiniMax News, February 12, 2026. https://www.minimax.io/news/minimax-m25

[52] "MiniMax M2.7: Early Echoes of Self-Evolution." MiniMax News, March 18, 2026. https://www.minimax.io/news/minimax-m27-en

[53] "LICENSE" (MiniMax-M2.7 non-commercial license). MiniMaxAI/MiniMax-M2.7, Hugging Face. https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE

[54] "MiniMax Just Open Sourced MiniMax M2.7: A Self-Evolving Agent Model that Scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2." MarkTechPost, April 12, 2026. https://www.marktechpost.com/2026/04/12/minimax-just-open-sourced-minimax-m2-7-a-self-evolving-agent-model-that-scores-56-22-on-swe-pro-and-57-0-on-terminal-bench-2/

[55] "MiniMax M3: Frontier Coding, 1M Context, Native Multimodality, All in One Model." MiniMax blog, June 1, 2026. https://www.minimax.io/blog/minimax-m3

[56] MiniMaxAI model cards and licenses on Hugging Face (MiniMax-M3, MiniMax-M1). https://huggingface.co/MiniMaxAI/MiniMax-M3

[57] Xunhao Lai et al. "MiniMax Sparse Attention." arXiv:2606.13392, June 2026. https://arxiv.org/abs/2606.13392

[58] "MiniMax Speech 02: Pioneering a New Era of AI Speech Generation." MiniMax News, 2025. https://www.minimax.io/news/minimax-speech-02

[59] "Music generation model Music-01." MiniMax News, August 31, 2024. https://www.minimax.io/news/music-01

[60] "3 billion interactions with the world." MiniMax News (Partner Day), August 31, 2024. https://www.minimax.io/news/3-billion-interactions

[61] "Pay as You Go" pricing. MiniMax API Docs. https://platform.minimax.io/docs/guides/pricing-paygo

[62] "Models." MiniMax API Docs. https://platform.minimax.io/docs/guides/models-intro

[63] "Disney, Warner Bros. Discovery, NBCU Sue Chinese AI Company MiniMax, Alleging It 'Pirates and Plunders' Studios' Copyrighted Works on 'Massive Scale'." Variety, September 16, 2025. https://variety.com/2025/digital/news/disney-warner-bros-discovery-nbcu-lawsuit-minimax-chinese-ai-company-1236520395/

[64] "Disney Enterprises, Inc. v. Minimax." Loeb & Loeb, IP/Entertainment Case Law Updates, May 2026. https://www.loeb.com/en/insights/publications/2026/06/disney-enterprises-inc-v-minimax

[65] "Exclusive: Major UK broadcasters investigate AI video generator." Broadcast, December 16, 2024. https://www.broadcastnow.co.uk/tech-innovation/exclusive-major-uk-broadcasters-investigate-ai-video-generator/5200038.article

[66] "MiniMax用爱奇艺视频'投喂'海螺？视频平台开打版权战" (Did MiniMax feed iQiyi videos to Hailuo? Video platforms start a copyright war). Jiemian (界面新闻), February 5, 2025. https://www.jiemian.com/article/12312654.html

[67] "Jensen Huang lauds China's AI models as Nvidia gears up to resume chip exports." CNBC, July 16, 2025. https://www.cnbc.com/2025/07/16/jensen-huang-china-ai-models-nvidia-gears-up-to-resume-chip-exports-.html

[68] "MiniMax创始人闫俊杰：未来全球只会剩下5家大模型企业" (MiniMax founder Yan Junjie: only five large-model companies will remain worldwide). The Paper (澎湃新闻), July 4, 2024. https://www.thepaper.cn/newsDetail_forward_27962886

[69] "No commercial use allowed in License?" MiniMaxAI/MiniMax-M2.7 community discussion, Hugging Face, April 12, 2026. https://huggingface.co/MiniMaxAI/MiniMax-M2.7/discussions/6

[70] "Q&A About License." MiniMaxAI/MiniMax-H3 repository, Hugging Face. https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/QA-about-License.md

[71] "与世界的30亿次交互" (3 billion interactions with the world). MiniMax News (Chinese), August 31, 2024. https://www.minimaxi.com/news/%E4%B8%8E%E4%B8%96%E7%95%8C%E7%9A%8430%E4%BA%BF%E6%AC%A1%E4%BA%A4%E4%BA%92

[72] "OpenAI sets two-stage Sora shutdown with app closing April 2026 and API following in September." The Decoder, March 28, 2026. https://the-decoder.com/openai-sets-two-stage-sora-shutdown-with-app-closing-april-2026-and-api-following-in-september/

## External links

- [Official Website](https://www.minimaxi.com)

- [MiniMax Open Platform](https://platform.minimax.io)

- [MiniMax on GitHub](https://github.com/MiniMax-AI)

- [MiniMax on Hugging Face](https://huggingface.co/MiniMaxAI)

- [MiniMax-01 Technical Paper (arXiv)](https://arxiv.org/abs/2501.08313)

- [MiniMax-M1 Technical Paper (arXiv)](https://arxiv.org/abs/2506.13585)