MiniMax: $4B Shanghai AI Lab Behind M2.7 & Hailuo (2026)

MiniMax is a Shanghai AI lab founded in 2021, now public on HKEX (0100.HK) since January 2026. It builds the open-weight MiniMax M2.7 model and Hailuo video AI.

MiniMax is a Shanghai-based AI company founded in December 2021, now listed on the Hong Kong Stock Exchange (0100.HK) after a $618 million IPO in January 2026 at a market cap above HK$70 billion. Its flagship products are the MiniMax M-series LLMs (M2.7, M3) for agentic coding, the Hailuo video generator, and consumer apps like Talkie, serving 236 million users and 214,000 enterprise customers across 200+ countries with $79 million 2025 revenue.

MiniMax is a Shanghai AI lab founded in December 2021 by Yan Junjie, Yang Bin, and Zhou Yucong, former SenseTime engineers. It went public on the Hong Kong Stock Exchange (0100.HK) in January 2026, raising about $618 million. MiniMax builds the open-weight MiniMax M2.7 agentic coding model, priced at $0.30 per 1M input tokens, and the Hailuo video generation series used by over 236 million users.

Founded: 2021 · HQ: Shanghai, China · Team: 150-400 · CEO: Yan Junjie · Funding: ~$1.15B private funding raised pre-IPO (lead: Shanghai STVC Group, $300M Series B extension at $4B valuation, Jul 2025); IPO'd on HKEX Jan 2026 raising ~$618M · Valuation: Market cap above HK$70B (~$9B) after Jan 9, 2026 HKEX debut (0100.HK)

About MiniMax

MiniMax was founded in December 2021 in Shanghai by Yan Junjie, Yang Bin, and Zhou Yucong, all of whom previously worked at SenseTime, one of China's largest computer vision companies. The founders left to build broad foundation models and consumer AI products under a single roof rather than narrow research output, framing the company's mission as helping to "co-create intelligence with everyone" on the path to artificial general intelligence. MiniMax Group Inc. is incorporated for its Hong Kong listing, with operating headquarters in Shanghai. The company's flagship product line is the M-series of large language models. MiniMax M2 and M2.1 were updated in the fourth quarter of 2025, M2.5 launched in February 2026, and MiniMax-M2.7 followed on March 18, 2026 as a 230-billion-parameter Mixture-of-Experts model with 10 billion active parameters, 256 experts (8 active per token), a 204,800-token context window, and a 131,072-token max output. M2.7 shipped as open weights in April 2026 under a modified MIT license. On June 1, 2026, MiniMax released M3, introducing a new MiniMax Sparse Attention (MSA) architecture with a 1-million-token context window, native multimodal input, and claims of 15.6x faster decoding than M2. In the last six months MiniMax has shipped on an unusually fast cadence: M2.5 in February 2026, M2.7 in March 2026 (open-weighted in April), and M3 on June 1, 2026. M2.7 scores 56.2% on SWE-Bench Pro (versus 32.6% for M2.1) and 57.0% on Terminal-Bench 2, and the company describes it as the first MiniMax model that actively participates in its own training and evaluation loop, handling an estimated 30-50% of that work itself. Alongside the LLMs, MiniMax shipped Hailuo 2.3 and a faster Hailuo 2.3 Fast video model, plus MiniMax Speech 2.8 and MiniMax Music 2.5+. MiniMax has raised roughly $1.15 billion in private funding. The most recent private round was a $300 million Series B extension in July 2025 at a $4 billion valuation, led by Shanghai state-owned capital through the Shanghai Science and Technology Venture Capital (STVC) Group. Earlier investors include Alibaba Group, Tencent, HongShan Capital Group, Hillhouse Investment, IDG Capital, Yunqi Capital, Vitalbridge Capital, and GL Ventures. On January 9, 2026, MiniMax listed on the Hong Kong Stock Exchange under ticker 0100.HK, selling about 29.2 million shares at HK$151-165 each for gross proceeds near $618-619 million, with cornerstone backing from Alibaba, the Abu Dhabi Investment Authority, Boyu Capital, and Mirae Asset. Shares rose as much as 21% on debut, pushing market capitalization above HK$70 billion. MiniMax makes money through API consumption on its M-series and multimodal models (priced from $0.30 per 1M input tokens and $1.20 per 1M output tokens for M2.7), enterprise contracts, and consumer subscriptions inside Talkie and Hailuo AI. For full-year 2025, MiniMax reported revenue of $79.0 million, up 158.9% year over year, with gross profit of $20.1 million at a 25.4% gross margin, up from roughly $10 million ARR in 2024. Over 70% of 2025 revenue came from international markets, and by December 31, 2025 the company had served more than 236 million cumulative users and 214,000 enterprise customers across over 200 countries. CEO Yan Junjie, 36, led the company through its Hong Kong IPO and became a billionaire with an estimated net worth around $3.2 billion. Zhao Pengyu, who joined MiniMax in August 2023 as an NLP researcher, was appointed an Executive Director in June 2025 and now leads large language model research and engineering. Reported headcount varies by source: PitchBook's 2026 data lists around 385 employees, while earlier 2025 trackers cited 150-200+. The core research and engineering team is based in Shanghai, with finance and investor relations functions added in Hong Kong after the listing. MiniMax's research output includes published work on Lightning Attention and test-time compute scaling (the MiniMax-01 and MiniMax-M1 papers), and its GitHub organization hosts over two dozen public repositories covering model weights, the Mini-Agent framework, and a multimodal generation CLI. Unlike Anthropic or OpenAI, MiniMax has not published a formal responsible scaling policy, named external red-teaming partners, or released detailed system cards; training data composition for the M-series is largely undisclosed. The company's safety communication runs mainly through its terms of service and a data-handling FAQ stating that customer API inputs are not used for training unless a customer opts in. Competitively, MiniMax sits among a cluster of Chinese frontier labs: DeepSeek and Alibaba's Qwen team on open-weight coding models, Moonshot AI's Kimi on long-context agents, and Zhipu AI's GLM on enterprise distribution. MiniMax-M2.7 is explicitly benchmarked against Claude Opus 4.6, scoring within 1.2 points on PinchBench (86.2% vs Opus) at a fraction of the price, though it trails on raw reasoning benchmarks like GPQA Diamond and is notably slower (35-46 tokens/sec versus a 60-95 t/s median for similarly priced models). In video, Hailuo 2.3 competes with Runway, Pika, and Kuaishou's Kling. As a Shanghai-headquartered company with state-owned capital among its investors and a Hong Kong listing, MiniMax operates under China's Generative AI Measures from the Cyberspace Administration of China alongside any GDPR obligations for its EU-facing products. No SOC 2 Type II report, ISO 27001 certificate, or HIPAA-eligible tier was found in public sources as of mid-2026, which limits its appeal to regulated US and EU enterprise buyers compared with Western frontier labs. Looking ahead, MiniMax's June 2026 M3 release (1M context, native multimodality, planned open-source) signals the company's strategy of combining aggressive open-weight distribution with rapid international consumer growth to fund continued AGI-focused research.

Mission

Co-create intelligence with everyone, advancing toward artificial general intelligence (AGI).

Products

Links

Website · GitHub · Twitter · Blog · Docs

Frequently Asked Questions

What is MiniMax and what do they build?

MiniMax is a Shanghai-based artificial intelligence company founded in December 2021 by Yan Junjie, Yang Bin, and Zhou Yucong, all former engineers at SenseTime. The company builds large language models under the M-series brand, including MiniMax M2, M2.1, M2.5, M2.7, and the June 2026 release M3, alongside the Hailuo video generation models, MiniMax Audio for speech and music, and consumer apps Talkie and MiniMax Agent. MiniMax-M2.7, released March 18, 2026, is a 230-billion-parameter Mixture-of-Experts model with only 10 billion active parameters, a 204,800-token context window, and shipped as open weights under a modified MIT license. It scores 56.2% on SWE-Bench Pro and 57.0% on Terminal-Bench 2, putting it among the strongest open agentic coding models at its price point. Developers can reach the M-series through MiniMax's own API platform (platform.minimax.io) or through Fireworks, Together.ai, OpenRouter, and several other inference providers. By the end of 2025, MiniMax had served more than 236 million users and 214,000 enterprise customers across over 200 countries. The company listed on the Hong Kong Stock Exchange under ticker 0100.HK in January 2026. MiniMax's overall position is as one of China's leading frontier AI labs, competing directly with DeepSeek, Alibaba's Qwen team, and Moonshot AI's Kimi models.

Who founded MiniMax and who is the CEO?

MiniMax was founded in December 2021 in Shanghai by Yan Junjie, Yang Bin, and Zhou Yucong. Yan Junjie and Zhou Yucong both previously worked at SenseTime, one of China's largest computer vision and AI companies, before building general-purpose foundation models together. Yan Junjie serves as Founder and CEO, and led the company through its Hong Kong IPO on January 9, 2026, becoming a billionaire with an estimated net worth around $3.2 billion at age 36. Zhao Pengyu, who joined MiniMax in August 2023 as a natural language processing researcher, was appointed an Executive Director in June 2025 and now leads large language model research and engineering. The founding thesis was to build broadly useful AI models and consumer products under one roof, reflected in the company's stated mission to 'co-create intelligence with everyone' on the path to AGI. There has been no public CEO change since founding; Yan Junjie remains in charge as of mid-2026. MiniMax is structured as MiniMax Group Inc., a holding company incorporated for its Hong Kong listing, with operations based in Shanghai. No separate long-term benefit trust or safety-focused governance body has been publicly disclosed, unlike some Western labs.

How much funding has MiniMax raised?

MiniMax raised approximately $1.15 billion in private funding before going public. A key milestone was a $300 million Series B extension closed in July 2025 at a $4 billion valuation, led by Shanghai state-owned capital through the Shanghai Science and Technology Venture Capital (STVC) Group. Earlier backers across MiniMax's funding history include Alibaba Group, Tencent, HongShan Capital Group (formerly Sequoia China), Hillhouse Investment, IDG Capital, Yunqi Capital, Vitalbridge Capital, and GL Ventures. On January 9, 2026, MiniMax completed its Hong Kong Stock Exchange IPO under ticker 0100.HK, selling around 29.2 million shares at HK$151 to HK$165 each for gross proceeds of roughly $618-619 million, with cornerstone investments from Alibaba, the Abu Dhabi Investment Authority, China's Boyu Capital, and Mirae Asset. The stock surged as much as 21% on debut, pushing market capitalization above HK$70 billion (roughly $9 billion). For fiscal year 2025, MiniMax reported total revenue of $79.0 million, up 158.9% year over year, with gross profit of $20.1 million at a 25.4% gross margin. The company has not announced any acquisitions; it has instead been a destination for AI talent in China's competitive labor market since 2022.

What products does MiniMax make?

MiniMax's core product line is the M-series of large language models: MiniMax M2 and M2.1 (updated in Q4 2025), M2.5 (February 2026), M2.7 (March 2026, open weights), and M3 (June 1, 2026, with a new MiniMax Sparse Attention architecture and 1-million-token context). These models target developers building coding agents and are priced from about $0.30 per 1 million input tokens and $1.20 per 1 million output tokens for M2.7 on MiniMax's own API. For video, MiniMax offers the Hailuo series, with Hailuo 2.3 and a faster Hailuo 2.3 Fast variant generating high-fidelity video clips that the company says cut batch content creation costs by up to 50%. MiniMax Audio covers text-to-speech and music generation, with Speech 2.8 and Music 2.5+ as the current versions. On the consumer side, Talkie is an AI companion and character chat app, MiniMax Agent is a general-purpose agent platform, and Hailuo AI is the consumer-facing video app; together these reached over 236 million cumulative users by the end of 2025. Enterprise and developer access runs through MiniMax's open platform at platform.minimax.io, plus third-party hosting via Fireworks, Together.ai, OpenRouter, and Novita. M2.7 and M3 are released as open weights (M2.7 under a modified MIT license restricting unauthorized commercial use), while earlier M-series and multimodal models remain proprietary, API-only.

Where is MiniMax headquartered and how big is the team?

MiniMax is headquartered in Shanghai, China, where its research and engineering teams are concentrated. Public headcount estimates vary by source and date: PitchBook's 2026 data lists around 385 employees, while earlier 2025 estimates from Wikipedia and other trackers put the figure above 200. The company has grown quickly alongside its revenue, which rose from about $10 million ARR in 2024 to roughly $70-79 million in 2025. MiniMax does not publicly break down headcount by office location, but its primary research hub remains Shanghai, with the Hong Kong listing adding investor relations and finance functions in Hong Kong. The company has not announced layoffs; instead it has been actively hiring for AGI-focused research and engineering roles through its careers page, including campus recruiting and internship programs. As a 2021-founded company that IPO'd within five years, MiniMax's team has scaled from a small founding group of former SenseTime researchers to a multi-hundred-person organization spanning model research, infrastructure, and international go-to-market for its 70%-plus international revenue base. No major restructuring or country-entry announcements were found for 2025-2026 beyond the Hong Kong listing itself. The company operates primarily in Mandarin and English across its product and developer documentation.

What is MiniMax's mission or research focus?

MiniMax's stated mission is to 'co-create intelligence with everyone,' framing its work around the long-term goal of artificial general intelligence (AGI) developed collaboratively rather than behind closed doors. In practice this shows up as MiniMax's unusually aggressive open-weight strategy: MiniMax-M2.7 (230B parameters, 10B active) was released with open weights on Hugging Face and ModelScope in April 2026, and the company has signaled that M3, launched June 1, 2026, will also be open-sourced. Research-wise, MiniMax has published technical work on Lightning Attention and test-time compute scaling in its MiniMax-01 and MiniMax-M1 papers on arXiv, and M2.7 is described by the company as its first model that actively participates in its own development cycle, handling an estimated 30-50% of its own training and evaluation work. The company's GitHub organization (MiniMax-AI) hosts over two dozen public repositories, including model weights, the Mini-Agent framework, and a CLI for text, image, video, speech, and music generation. Compared with Anthropic or OpenAI, MiniMax has not published a formal responsible scaling policy or named external red-teaming partners; its safety posture is communicated mainly through terms of service and a data-handling FAQ rather than detailed system cards. Training data composition for the M-series is largely undisclosed. The mission statement leans toward broad accessibility and commercial deployment speed, evidenced by the company's emphasis on price-performance over published safety research. MiniMax frames its AGI ambitions alongside consumer reach, citing its 236 million-user base as proof that frontier models can be deployed at scale outside narrow developer audiences.

Is MiniMax compliant with SOC 2, GDPR, HIPAA?

MiniMax states that its API platform uses TLS encryption for all traffic, role-based access controls, and regular third-party penetration testing, and describes its security program as SOC 2-aligned, though a specific SOC 2 Type II report or public trust center URL was not found during research. By default, customer inputs sent to the MiniMax API for inference are not stored or used to train MiniMax's models unless a customer explicitly opts in, according to the company's API privacy policy. MiniMax offers data residency controls that let enterprise customers choose processing regions, with infrastructure described as spanning North America, Europe, and Asia-Pacific. For GDPR, MiniMax's privacy policy addresses EU users, but no SCCs, DPA template, or dedicated EU representative were independently verified. No HIPAA-eligible tier, ISO 27001 certificate, or FedRAMP authorization was found in public sources as of mid-2026. Authentication on the developer platform supports OAuth 2.0, and organizations can enable multi-factor authentication for team accounts. As a Shanghai-based company with a Hong Kong listing, MiniMax is also subject to China's Generative AI Measures from the Cyberspace Administration of China alongside any GDPR obligations for EU-facing products. Enterprises with strict compliance requirements should request current audit documentation directly from MiniMax's API support team before relying on any specific certification.

Who are MiniMax's main competitors?

MiniMax's closest competitors are other Chinese frontier model labs: DeepSeek, Alibaba's Qwen team, Zhipu AI (GLM), and Moonshot AI (Kimi). Against DeepSeek, MiniMax competes directly on open-weight coding models; MiniMax-M2.7's 56.2% SWE-Bench Pro score and $0.30/$1.20 per-million-token pricing position it as a cheaper, if slower, alternative, since M2.7 outputs around 35-46 tokens per second versus a roughly 60-95 t/s median for similarly priced open models. Against Alibaba's Qwen, MiniMax has less enterprise cloud distribution in China since Qwen is bundled into Alibaba Cloud, but MiniMax counters with a larger international consumer footprint via Talkie and Hailuo. Against Moonshot AI's Kimi models, both companies focus heavily on long-context agentic use cases, with MiniMax differentiating on its self-evolving M2.7 training approach and broader multimodal lineup spanning video, speech, and music. On the Western side, MiniMax-M2.7 is explicitly benchmarked against Claude Opus 4.6, scoring within 1.2 points on PinchBench (86.2% vs Opus) at a fraction of the price, though it trails frontier US and Chinese models on raw reasoning benchmarks like GPQA Diamond. In video generation, Hailuo 2.3 competes with Runway, Pika, and Kuaishou's Kling. One category where MiniMax has no strong public answer yet is enterprise compliance certification compared to Anthropic, OpenAI, or Google, which limits its appeal for regulated US and EU enterprise buyers. A newer entrant to watch is MiniMax's own M3 (June 2026), positioned to leapfrog M2.7 and rival open models with a 1-million-token context and native multimodality.