CtrlK
| MiniMax |
|---|
| 上海稀宇科技有限公司 (Shanghai Xiyu Technology Co., Ltd.) |
 |
| Type |
| Industry |
| Founded |
| Founders |
| Zhou Yucong (周雨丛) |
| Headquarters |
| Key people |
| Wei Wei (Legal Representative) |
| Parent |
| Products |
| Hailuo AI (海螺AI) |
| MiniMax Chat |
| Hai Luo (海螺) |
| ABAB series |
| MiniMax-01 series |
| MiniMax-M1 |
| MiniMax-M2 |
| Speech-02 |
| Music-01 |
| Revenue |
| Valuation |
| Employees |
| Website |
MiniMax (Chinese: 上海稀宇科技有限公司, Shanghai Xiyu Technology Co., Ltd.) is a publicly traded artificial intelligence (AI) company headquartered in Shanghai, China. Founded in December 2021 by former SenseTime executives, the company has been dubbed one of China's "Six Little Tigers" (六小虎) of AI by investors and media.[3] MiniMax develops large language models, multimodal AI systems, and consumer-facing AI applications spanning text, video, speech, and music generation. The company's name is derived from the minimax algorithm in game theory, which seeks to minimize potential losses in a worst-case scenario while maximizing potential gains.
As of December 31, 2025, MiniMax's products have served over 236 million individual users across more than 200 countries and regions, and more than 214,000 enterprises and developers across over 100 countries and regions.[1] The company processes roughly 3 billion daily AI interactions and 30 trillion tokens according to founder Yan Junjie.[4] On January 9, 2026, MiniMax went public on the Hong Kong Stock Exchange under stock code 9660, with shares more than doubling on their first day of trading.[2]
MiniMax was founded in December 2021 by several computer vision veterans from SenseTime:[5][6]
Yan Junjie (闫俊杰) - Former Vice President at SenseTime, responsible for building deep learning toolchains and underlying algorithms, now serves as CEO. Born in 1989 in a small town in Henan province, Yan taught himself advanced calculus in high school. He graduated from Southeast University with a degree in mathematics in 2006, obtained a PhD from the Institute of Automation at the Chinese Academy of Sciences in 2015, and completed postdoctoral research at Tsinghua University's Computer Science Department with a focus on deep learning and computer vision. Before founding MiniMax, he spent more than six years at SenseTime, where he served as deputy head of its research institute overseeing development of deep-learning technologies, distributed computing, inference, and general intelligence systems.[7]
Yang Bin (杨斌) - Previously co-founded the research center for Uber Advanced Technologies Group and worked at Waabi
Zhou Yucong (周雨丛) - Former head of algorithms R&D at SenseTime
When it first started out, MiniMax received early-stage funding from MiHoYo, the Chinese video game developer known for Genshin Impact.[8]
MiniMax has successfully raised significant funding through multiple rounds, making it one of the best-funded AI startups in China:
| Date | Round | Amount | Lead Investor(s) | Post-Money Valuation | Reference |
|---|---|---|---|---|---|
| 2022 | Seed/Early | Undisclosed | MiHoYo | Undisclosed | [8] |
| June 2023 | Series A | $250+ million | Tencent-backed entity and others | ~$1.2 billion | [8] |
| March 2024 | Series B | $600 million | Alibaba Group | ~$2.5 billion | [9][10] |
| 2024-2025 | Additional rounds | ~$300 million | Multiple (reported) | >$4 billion | [11] |
| Total Raised (Pre-IPO) | $1.15+ billion |
Other investors include Hillhouse Investment, HongShan Capital Group (formerly Sequoia China), IDG Capital, Vitalbridge Capital, GL Ventures, Tencent Holdings, and the Shanghai State-owned Assets Supervision and Administration Commission.[9][10]
On July 16, 2025, Reuters reported that MiniMax had confidentially filed for an initial public offering (IPO) on the Hong Kong Stock Exchange.[12] The company became one of the fastest AI startups globally to go from founding to public listing, accomplishing it in roughly four years.
MiniMax priced its IPO at HK$165 per share, at the top of its marketed range, selling 25.4 million shares and raising approximately HK$4.2 billion (US$538 million). After the greenshoe (over-allotment) option was exercised, total proceeds reached HK$4.8 billion (approximately US$620 million). CICC and UBS served as sponsors for the listing.[2][13]
On its debut trading day of January 9, 2026, MiniMax shares surged 109%, closing at HK$345 per share. This gave the company a market capitalization exceeding HK$100 billion (roughly US$13 billion), far above the pre-IPO valuation target of around US$4 billion. The IPO was notable for landing just one day after fellow Chinese AI startup Zhipu AI debuted on the same exchange.[2][14]
In the weeks following the IPO, MiniMax stock continued to climb. By March 2026, the share price had surged past HK$1,300, pushing market capitalization above HK$300 billion. The stock received an additional boost in March 2026 when NVIDIA CEO Jensen Huang praised Chinese AI companies including MiniMax, causing shares to jump over 20% in a single session.[15]
MiniMax's revenue has grown rapidly, though the company continues to invest heavily and has not yet reached profitability:
| 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% |
More than 70% of MiniMax's 2025 revenue came from international markets, reflecting the global popularity of products like Talkie and Hailuo AI.[1] Sales and distribution expenses fell 40.3% from US$87.0 million to US$51.9 million as the company improved its go-to-market efficiency.
MiniMax's first product was Glow, launched in October 2022, which allowed users to create virtual characters, give them background stories and chat with them about various topics. Within four months of launch, the app had over 5 million users. Due to filing issues, Glow was terminated in March 2023.[16]
Glow was relaunched under two new brands:
Talkie for international markets (June 2023)
Xing Ye (星野) for the Chinese market (September 2023)
Talkie is a companion AI chatbot app that enables users to create and converse with virtual characters based on fiction or real people using generative AI technology. By mid-2024, Talkie had emerged as one of the most popular AI entertainment apps worldwide, directly competing with Character.ai.
For the first half of 2024, Talkie ranked as the fourth most-downloaded AI app in the United States, ahead of Google-backed Character.ai which ranked tenth. More than half of its 11 million monthly active users were located in the U.S. Other popular markets included the Philippines, the United Kingdom, and Canada.[17] Globally, Talkie recorded 17 million downloads in the first eight months of 2024. The app's user base skews young, with those aged 18 to 35 accounting for more than 70% of users.[18]
The app features AI avatars of public figures, including Donald Trump, Taylor Swift, Elon Musk, and LeBron James.[19]
Talkie was pulled from Apple's App Store in December 2024 for unspecified "technical" reasons but later returned.[19] Despite this disruption, Talkie's global traction has been a key driver of MiniMax's international revenue, which accounts for over 70% of the company's total.
Launched in March 2024, Hailuo AI is a multimodal large language model consumer platform that provides AI text, video, and music-generating features. Hailuo AI has become one of the most widely used AI video generation platforms globally, competing with Sora from OpenAI, Runway Gen-3, and Kling from Kuaishou. Key releases include:
| Model | Type | Release | Key Capabilities |
|---|---|---|---|
| Video-01 | Text-to-video | September 2024 | 720p at 25fps, up to 6 seconds, cinematic camera movement, MP4 output (H.265 codec) |
| T2V-01-Director | Text-to-video (director mode) | January 2025 | Professional filmmaking tools for controlled video generation |
| I2V-01-Director | Image-to-video (director mode) | January 2025 | Professional-grade image-to-video with cinematic controls |
| I2V-01-Live | Image-to-video (animation) | 2025 | Smooth, vivid animation optimized for 2D illustrations |
| S2V-01 | Subject-to-video | 2025 | Character-consistent video from a single reference image; maintains facial identity across angles |
| Hailuo 02 | Next-gen video | June 2025 | Native 1080p, Noise-aware Compute Redistribution (NCR) architecture, 3x parameters, 4x training data vs Video-01, extreme physics mastery |
| Hailuo 2.3 | Enhanced video | Late 2025 | Improved physical actions, stylization, character micro-expressions, anime/illustration support, 50% cost reduction with 2.3 Fast variant |
Hailuo 02 introduced a core framework based on Noise-aware Compute Redistribution (NCR), which boosts training and inference efficiency by 2.5 times at comparable parameter scales. It ships in three resolution/duration variants: 768p for 6 seconds, 768p for 10 seconds, and 1080p for 6 seconds.[20]
In 2025, Hailuo AI introduced a "Media Agent" feature that evolved from the earlier Video Agent. The Media Agent supports comprehensive multimodal creation and automatically selects the appropriate models, offering a "one-click video generation" workflow that handles the entire pipeline without manual editing.[21]
Hai Luo is an AI-powered productivity and copilot tool, similar to Microsoft Copilot. It can perform tasks like document summarization, content creation, and code generation. It supports processing long documents up to 200,000 words.[22]
MiniMax has developed a succession of increasingly capable AI models across text, vision, speech, and music:
The ABAB series represents MiniMax's earlier generation of large language models:
| Model | Type | Release | Key Features |
|---|---|---|---|
| ABAB 5 | LLM | August 2023 | Foundation for AI chatbot and productivity tools |
| ABAB 5.5 | LLM | 2023 | Context window of 16,384 tokens (ABAB 5.5s: 8,192 tokens) |
| ABAB 6 | Mixture of experts (MoE) | January 2024 | 1 trillion parameters (claimed), comparable to GPT-4 on certain benchmarks, 32,768 token context[23] |
| ABAB 6.5 | Mixture of experts | April 17, 2024 | Trillion-parameter MoE, 200K token context window; ABAB 6.5s processes ~30,000 words per second[24] |
The MiniMax-01 series, released in January 2025, introduced a novel hybrid architecture combining Lightning Attention, Softmax Attention, and Mixture-of-Experts (MoE). The series was published as an open-source release with weights and code on Hugging Face and GitHub.[25]
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 achieves linear complexity by dividing the attention calculation into two distinct components: intra-block computations (using left-product attention) and inter-block computations (using right-product attention). This design maintains an information capacity of O(d^2/h), which is larger than standard softmax attention's capacity of O(d), where d is the model dimension and h is the number of attention heads.[25]
To train these models efficiently, MiniMax developed several parallelism strategies including Linear Attention Sequence Parallelism Plus (LASP+), varlen ring attention, and Expert Tensor Parallel (ETP).[25]
| 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[25] |
| MiniMax-VL-01 | Multimodal LLM | 456 billion total | 1 million+ | Visual and text understanding capabilities[25] |
MiniMax-Text-01 matches the performance of state-of-the-art models like GPT-4o and Claude 3.5 Sonnet on standard benchmarks while offering a context window 20 to 32 times longer.[25]
MiniMax-M1, released in June 2025, is the world's first open-weight, large-scale hybrid-attention reasoning model. It combines the Lightning Attention architecture from MiniMax-01 with reinforcement learning-based reasoning capabilities, supporting a 1-million-token context window.[26]
The entire reinforcement learning phase for M1 used only 512 NVIDIA H800 GPUs for three weeks, with a rental cost of just $534,700. The model requires only 30% of the computing power needed by rival DeepSeek's R1 model when performing deep reasoning tasks, representing a significant breakthrough in computational efficiency.[26]
MiniMax-M2, released in late 2025, is a compact and cost-effective MoE model with 230 billion total parameters and 10 billion active parameters per token, released under the MIT license for both commercial and non-commercial use.[27]
| Model | Release | Key Features |
|---|---|---|
| MiniMax-M2 | Late 2025 | 230B total / 10B active parameters, top-tier agentic and tool-calling performance, MIT license[27] |
| MiniMax-M2.1 | Early 2026 | Optimized for robustness in coding, tool use, instruction following, and long-horizon planning |
| MiniMax-M2.5 | February 2026 | Polyglot code mastery, enhanced multilingual capabilities |
| MiniMax-M2.7 | March 2026 | Proprietary "self-evolving" model, performs 30-50% of RL research workflow automatically[28] |
MiniMax-M2 achieved the highest scores among open-weight models for real-world agentic and tool-calling tasks at the time of release, placing it at or near the level of top proprietary systems like GPT-5 and Claude Sonnet 4.5.[27]
| Model | Release | Key Features |
|---|---|---|
| Speech-01 | 2024 | Initial text-to-speech offering |
| Speech-02 | April 2025 | 32 languages, 99% vocal similarity from 10 seconds of reference audio; two variants: Speech-02-HD (high-fidelity for voiceovers/audiobooks) and Speech-02-Turbo (ultra-low latency for real-time applications)[29] |
Speech-02 uses a learnable timbre extractor (Learnable Speaker Encoder) integrated with an AR Transformer architecture. It supports English variants (US, UK, Australian, Indian), Asian languages (Mandarin, Cantonese, Japanese, Korean, Vietnamese, Indonesian), European languages (French, German, Spanish, Portuguese, Turkish, Russian, Ukrainian), and more. It earned top honors on both the Artificial Analysis Speech Arena and Hugging Face TTS Arena benchmarks.[29]
Music-01, introduced in August 2024, is an AI music generation model that simultaneously generates both accompaniment and vocals. Users can upload reference music for the model to learn rhythm and style, then input lyrics to receive a new AI-generated piece. It supports classical, pop, rock, electronic, and dozens of additional styles. The current maximum output length is 60 seconds with up to 400 characters of lyrics.[30]
| Model Name | Type | Release Date | Key Features |
|---|---|---|---|
| ABAB 5 | LLM | August 2023 | Foundation chatbot model |
| ABAB 5.5 | LLM | 2023 | 16,384 token context |
| ABAB 6 | MoE LLM | January 2024 | Trillion parameters, GPT-4 comparable |
| ABAB 6.5 | MoE LLM | April 2024 | Trillion parameters, 200K context |
| Music-01 | Music generation | August 2024 | Simultaneous vocals and accompaniment |
| Video-01 | Text-to-video | September 2024 | 720p, 6 seconds, cinematic quality |
| MiniMax-Text-01 | MoE LLM | January 2025 | 456B params, 4M token context, Lightning Attention |
| MiniMax-VL-01 | Multimodal LLM | January 2025 | Vision + text understanding |
| Speech-02 | Text-to-speech | April 2025 | 32 languages, 99% vocal similarity |
| Hailuo 02 | Video generation | June 2025 | Native 1080p, NCR architecture |
| MiniMax-M1 | Reasoning LLM | June 2025 | Hybrid-attention reasoning, 30% compute of DeepSeek R1 |
| MiniMax-M2 | Open-source LLM | Late 2025 | 230B/10B active, MIT license, top agentic performance |
| S2V-01 | Subject-to-video | 2025 | Character-consistent video from single image |
| Hailuo 2.3 | Enhanced video | Late 2025 | Improved physics, stylization, micro-expressions |
| MiniMax-M2.5 | Open-source LLM | February 2026 | Polyglot code mastery |
| MiniMax-M2.7 | Proprietary LLM | March 2026 | Self-evolving, automated RL research |
MiniMax provides enterprise API services through its Open Platform (platform.minimax.io), serving over 214,000 enterprises and developers globally as of December 2025. The platform offers access to text, vision, video, speech, and music generation models.
The company offers competitive pricing for its API services:
| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|
| MiniMax-M2 | $0.255 | $1.00 |
| MiniMax-M2.5 | $0.20 | ~$0.80 |
| MiniMax-Text-01 (up to 200K context) | $0.40 | $1.30 |
| MiniMax-Text-01 (up to 1M context) | $1.30 | $4.20 |
The platform is supported by a multi-cloud GPU infrastructure of 10,000+ GPUs, serving billions of API calls per day. Enterprise features include 99.99% uptime SLA, automatic failover, load balancing, and 24/7 monitoring.[31]
MiniMax operates through its parent company, Shanghai Xiyu Jizhi Technology Co., Ltd. (上海稀宇极智科技有限公司), which was established in 2021. The company is a subsidiary of MiniMax HONGKONG Limited with 100% ownership. Following the January 2026 IPO, MiniMax trades publicly on the Hong Kong Stock Exchange under stock code 9660.[32]
MiniMax is one of China's "Six Little Tigers" (六小虎), a designation for the leading Chinese AI startups focused on large language models and multimodal AI. The Six Little Tigers represent a new generation of Chinese AI companies that emerged after the launch of ChatGPT in late 2022, distinct from the earlier "Four Little Dragons" (SenseTime, Megvii, CloudWalk Technology, and Yitu Technology) that primarily focused on computer vision. The six companies are:[3][33]
| Company | Focus | Notable Products | Valuation (approx.) |
|---|---|---|---|
| MiniMax | Multimodal AI, consumer apps | Talkie, Hailuo AI, ABAB/M-series models | >US$13 billion (post-IPO) |
| Zhipu AI | General LLMs, enterprise AI | GLM series, ChatGLM | >US$3 billion |
| Moonshot AI | Long-context AI | Kimi chatbot | ~US$3 billion |
| Baichuan Intelligence | General LLMs | Baichuan series | >US$2 billion |
| 01.AI | Cost-efficient LLMs | Yi series | ~US$1 billion |
| StepFun (阶跃星辰) | Multimodal AI | Step series | >US$2 billion |
The term evolved from the earlier "Four Little Tigers" designation used in 2024, which included MiniMax, Zhipu AI, Moonshot AI, and Baichuan Intelligence.[34]
MiniMax competes across multiple AI segments:
In July 2025, NVIDIA CEO Jensen Huang met with MiniMax founder Yan Junjie in Beijing and publicly praised China's AI innovation, stating that models from companies like "DeepSeek, Alibaba, Tencent, MiniMax, and Baidu" are "world class" and "have spurred AI developments worldwide."[15] Following the meeting, Huang commented that developers and entrepreneurs in China are driving AI technology innovation at an astonishing speed.
In 2024, MiniMax founder Yan Junjie predicted that "in the future, there will only be five large model companies left in the world," noting that the ratio of the AI market between internet companies and AI startups may reach 9:1, more extreme than the 6:4 ratio seen in the transition from internet to mobile internet.[35]
MiniMax's primary technical contribution to the AI field is Lightning Attention, a linear-complexity attention mechanism that enables efficient processing of extremely long sequences. Published alongside the MiniMax-01 release (arXiv: 2501.08313), Lightning Attention works by splitting the attention computation into intra-block and inter-block operations:[25]
This hybrid approach allows models to handle context windows of 1 million tokens during training and extrapolate to 4 million tokens during inference at affordable cost. MiniMax combined Lightning Attention with standard softmax attention layers in a ratio that balances expressiveness and efficiency.[25]
In January 2025, MiniMax released its MiniMax-01 model family, including weights and code. CEO Yan Junjie admitted he would have gone open-source from day one if he could choose again, as the decision invites developer collaboration while showcasing technical capabilities. The MiniMax-M2 series continued this open-source strategy with an MIT license, allowing unrestricted commercial use.[25][27]
MiniMax has developed several training optimization techniques for distributed computing at scale:[25]
In September 2025, Disney, Warner Bros. Discovery, and NBCUniversal filed a joint copyright infringement lawsuit against MiniMax and its parent company Shanghai Xiyu Jizhi Technology Co. Ltd., alleging "willful and brazen" copyright infringement through its Hailuo AI service. The studios alleged that MiniMax markets Hailuo AI as a "Hollywood studio in your pocket" and built its business from intellectual property stolen from Hollywood studios.[36][37]
In December 2024, Broadcast magazine reported that Hailuo AI can reproduce the logos of British television channels Channel 4, Channel 5, and ITV in its AI-generated videos.[38]
MiniMax is reportedly being sued by iQiyi, a Chinese video streaming service that alleges MiniMax illicitly trained on iQiyi's copyrighted recordings.[39]
The sudden rise of DeepSeek in early 2025 disrupted the strategies of several of the Six Little Tigers, including MiniMax. Several startups had to quickly adjust their businesses and in some cases downsize teams. However, MiniMax weathered this disruption better than some peers by doubling down on consumer applications and multimodal AI rather than competing solely on base model performance.[33]