MiniMax AI: The Company Behind MaxClaw
MiniMax is the AI company that built MaxClaw, the cloud-hosted AI agent deployed through the MiniMax Agent platform. Founded in 2021, MiniMax is one of China's "Six AI Tigers" and a pioneer in full-stack AI development — building proprietary foundation models, inventing Lightning Attention, and scaling consumer products to over 212 million users globally. This page provides a comprehensive profile of MiniMax's technology, products, business strategy, and competitive positioning.
Company Overview
MiniMax was founded in 2021 with an ambition that set it apart from most AI startups of that era: build the entire stack. Rather than focusing exclusively on model research or application development, MiniMax pursued a "full-stack" approach — developing proprietary foundation models from the ground up while simultaneously building massive consumer-facing products on top of them.
This strategy placed MiniMax among China's "Six AI Tigers," the six most influential AI companies in the country. The group also includes Zhipu AI, Moonshot AI (creator of Kimi), Baichuan, StepFun, and 01.AI. Among these, MiniMax distinguishes itself through its consumer-centric focus and its willingness to ship products at global scale.
By late 2025, MiniMax had accumulated over 212 million individual users globally — a figure driven primarily by the success of its consumer products Xingye (Talkie) and Hailuo AI. This user base, combined with its proprietary model infrastructure, gives MiniMax a data flywheel that few AI companies outside of the largest incumbents can match.
Core Technological Innovations
MiniMax's technical contributions center on three innovations that address fundamental bottlenecks in Transformer-based AI models: Lightning Attention, a hybrid attention architecture, and Mixture-of-Experts scaling.
Lightning Attention
Standard Transformer models rely on the SoftMax attention mechanism, which scales quadratically with sequence length. This means that as context windows grow, the computational cost grows exponentially — making million-token context windows prohibitively expensive with traditional approaches.
MiniMax's answer is Lightning Attention, a linear attention mechanism that replaces the quadratic complexity of SoftMax with linear scaling. This allows context windows to grow to millions of tokens without the corresponding explosion in compute cost. Lightning Attention is the core enabling technology behind MiniMax's ability to offer 4-million token context windows.
Hybrid Architecture
MiniMax does not use Lightning Attention exclusively. The MiniMax-01 and M1 series models employ a hybrid design that interleaves 7 Lightning Attention layers with 1 traditional SoftMax attention layer. This ratio provides the efficiency benefits of linear attention for the majority of computation while preserving the expressiveness of SoftMax attention where it matters most.
This hybrid approach is what enabled MiniMax-01 to support a 4-million token context window — surpassing the context lengths of GPT-4o and Claude 3.5 at the time of its release.
Mixture-of-Experts (MoE)
The MiniMax M2.5 model — the foundation model that powers MaxClaw — uses a Mixture-of-Experts architecture. M2.5 contains 229 billion total parameters, but only approximately 10 billion are activated for any given token. This sparse activation pattern means that M2.5 delivers intelligence comparable to much larger dense models while requiring a fraction of the compute per inference.
For MaxClaw users, this translates to fast inference speeds, lower per-token costs (1/7 to 1/20 of comparable models like Claude 3.5 Sonnet), and strong performance across code generation, multi-step tool calling, and logical reasoning.
Product Ecosystem
MiniMax operates one of the broadest product portfolios among AI-native companies. Rather than focusing on a single application, MiniMax ships multiple consumer-grade products — each powered by its proprietary foundation models and each targeting a distinct use case.
| Product | Category | Key Metric | Highlight |
|---|---|---|---|
| Hailuo AI | Multimodal Assistant | 5.6M+ MAU | Video-01 engine: 720p video generation at 25fps |
| Xingye (Talkie) | AI Character Interaction | 200M+ users, 200 countries | Persistent memory, voice interaction, global reach |
| MiniMax Agent | Cloud Agent Platform | Hosts MaxClaw | One-click deployment, multi-platform integration |
| Video-01 | Text-to-Video | 720p / 25fps | Competes with Sora and Kling |
Hailuo AI
Hailuo AI is MiniMax's multimodal assistant, serving over 5.6 million monthly active users. Its standout capability is Video-01, a text-to-video generation engine that produces high-fidelity video at 720p resolution and 25 frames per second. Video-01 competes directly with OpenAI's Sora and Kuaishou's Kling in the emerging text-to-video space, and is the flagship technology of Hailuo AI's video and research applications.
Xingye (Talkie)
Xingye, known internationally as Talkie, is an AI character interaction platform with persistent memory and voice capabilities. Talkie has reached over 200 million users across 200 countries, making it one of the most successful AI-native consumer products worldwide. Users create and interact with AI characters that remember past conversations, adapt to individual preferences, and communicate through natural voice interaction.
MiniMax Agent and MaxClaw
MiniMax Agent is the cloud-managed agent platform that hosts MaxClaw. It provides the infrastructure for deploying autonomous AI agents with one-click setup, persistent long-term memory, and native integration with Telegram, Discord, and Slack. MaxClaw, as the flagship agent on this platform, inherits the full OpenClaw tool ecosystem and is powered by the MiniMax M2.5 model for cost-efficient, high-performance agentic tasks.
Business Strategy and Funding
MiniMax's business trajectory represents one of the fastest ascents in the global AI industry — from founding to public listing in under five years.
Company Founded
MiniMax established with a full-stack AI vision: proprietary foundation models combined with consumer-scale applications.
$600M Funding Round
Alibaba led a $600 million funding round, valuing MiniMax at $2.5 billion. Other backers include Tencent and HongShan (formerly Sequoia China).
212M+ Global Users
MiniMax surpassed 212 million individual users globally, driven by the international expansion of Talkie and the growth of Hailuo AI.
IPO on Hong Kong Stock Exchange
MiniMax went public on the HKEX, with its stock price surging over 100% on its debut. MiniMax became one of the first Chinese AI unicorns to go public.
The backing by Alibaba, Tencent, and HongShan (Sequoia China) provided MiniMax with both capital and strategic distribution advantages. Alibaba's cloud infrastructure and Tencent's social platform reach complement MiniMax's consumer-focused strategy, while HongShan's track record in scaling Chinese technology companies provided operational guidance during the rapid growth phase.
The January 2026 IPO marked a milestone not just for MiniMax, but for the broader Chinese AI industry. With its stock soaring over 100% on debut day, MiniMax demonstrated that AI-native companies with strong consumer metrics and proprietary technology can command public market premiums — a signal that has influenced the IPO ambitions of other Six Tigers members.
The Six AI Tigers: Competitive Landscape
China's "Six AI Tigers" represent six distinct strategies for building an AI company. While all six are developing foundation models, their approaches to commercialization, architecture, and target markets diverge significantly. Here is how MiniMax positions itself relative to three key competitors.
MiniMax vs Zhipu AI
Zhipu AI focuses heavily on B2B and enterprise applications, building AI solutions for corporate customers and government institutions. MiniMax takes the opposite approach: consumer-centric products designed for mass adoption. While Zhipu AI monetizes through enterprise contracts, MiniMax monetizes through high-MAU consumer apps and platform services like MiniMax Agent.
MiniMax vs Moonshot AI (Kimi)
Moonshot AI's Kimi was a pioneer in bringing long-context capabilities to the public. However, MiniMax's Lightning Attention handles 4 million tokens more efficiently than traditional long-context approaches. Both companies ship consumer products, but Moonshot AI is more focused on the Kimi browser ecosystem, while MiniMax spans video generation, character interaction, and agent platforms.
MiniMax vs DeepSeek
DeepSeek has earned recognition for its focus on model efficiency and its commitment to open-source research. MiniMax shares the efficiency focus (via MoE and Lightning Attention) but differs in strategy: MiniMax builds polished, high-MAU consumer applications on top of its models, while DeepSeek primarily releases models and research for the community.
MiniMax's Unique Position
Among the Six Tigers, MiniMax is the only company that has simultaneously achieved: a 200M+ user consumer app (Talkie), a state-of-the-art video generation system (Video-01), an agent deployment platform (MiniMax Agent / MaxClaw), and a public listing on the HKEX. This breadth is its defining competitive advantage.
MaxClaw's Role in MiniMax's Strategy
MaxClaw represents MiniMax's entry into the AI agent market — a category that extends beyond chat-based assistants into autonomous, tool-using, always-on AI systems that operate within users' existing workflows.
Why MaxClaw Matters for MiniMax
MiniMax's existing products — Talkie for social interaction and Hailuo AI for multimodal tasks — serve large user bases but operate primarily within MiniMax's own platforms. MaxClaw, deployed through MiniMax Agent, extends MiniMax's AI capabilities into third-party platforms: Telegram, Discord, and Slack.
This is strategically significant because it allows MiniMax to reach users where they already spend time, rather than requiring them to visit a MiniMax property. Combined with the cost efficiency of the M2.5 model (1/7 to 1/20 of Claude 3.5 Sonnet), MaxClaw makes it economically viable to run continuous, high-frequency AI agents — opening use cases like automated monitoring, scheduled analysis, and persistent workflow automation that would be prohibitively expensive on other platforms.
MaxClaw is, in effect, the distribution channel for MiniMax's model intelligence into the broader agent ecosystem. It sits at the intersection of MiniMax's three core strengths: proprietary models (M2.5), cloud infrastructure (MiniMax Agent), and consumer product design (one-click deployment, persistent memory, persona customization).