MiniMax Hits 300M Users and Doubles ARR in 60 Days: The Anatomy of China's Most Successful AI Export
*Photo: A laptop displaying global connectivity. MiniMax's AI services now reach users in 200+ countries, with 73% of revenue generated outside China. Image: Unsplash*
Executive Summary
On May 28, 2026, MiniMax co-founder and COO Yunye Wei sat down with Bloomberg and dropped three numbers that reset the global AI industry's understanding of Chinese startups: 300 million users, annual recurring revenue that doubled in two months, and over 1 million enterprise and developer clients—a 5x increase in six months.
The company that most Western investors had never heard of has become the most successful Chinese AI export since TikTok. MiniMax's revenue model is starkly different from American AI companies: 73% of its $79 million in 2025 revenue came from international markets, with the United States and Singapore as the top two markets. Its products—Talkie (AI companion), Hailuo AI (video generation), and a developer API platform—have found product-market fit across 200 countries.
But the story is not unambiguously triumphant. MiniMax lost $250 million in 2025 on $79 million in revenue. The company filed for an A-share IPO on June 1, 2026, the same day it launched its newest flagship model, MiniMax M3. The question facing investors is whether this is the beginning of a global AI giant—or a cautionary tale about the cost of scaling at any price.
| Metric | 2024 | 2025 | H1 2026 (est.) | Change |
|---|---|---|---|---|
| Total users | ~200M | ~250M | 300M+ | +50% in 6 months |
| Annual revenue | $30.5M | $79.0M | ~$150M (ARR) | +159% YoY |
| International revenue share | 69.8% | 73% | ~75% | +3.2pp |
| Enterprise clients | ~20K | ~50K | 100万+ | 5x in 6 months |
| Gross margin | 12.2% | 25.4% | ~30% | +13.2pp |
| Net loss (adjusted) | $244M | $250M | ~$200M (annualized) | Flat |
| ARR (reported) | ~$60M | ~$150M (Feb) | ~$300M (May) | Doubled in 60 days |
*Table 1: MiniMax core financial and operational metrics. Revenue figures from company filings and Bloomberg interviews. ARR = Annual Recurring Revenue. User figures are cumulative registered users across all products.*
The 60-Day Revenue Doubling: What Triggered It?
From $150M to $300M ARR: February to May 2026
MiniMax's February 2026 ARR was approximately $150 million, disclosed by CEO Junjie Yan during an earnings call. By May 28, that number had doubled. Two factors drove the acceleration, and understanding them is essential to evaluating whether this growth is sustainable.
Factor 1: The M2.5 Model Release and OpenRouter Dominance
MiniMax released its M2.5 model in late 2025, positioning it as a high-performance, low-cost alternative to Claude and GPT-4. The pricing was aggressive: $0.30 per million input tokens, compared to Claude Opus at $5.00 and GPT-4 Turbo at $10.00. This 16x-33x price differential, combined with performance that reached 96% of Claude Opus on coding benchmarks, created a viral adoption curve among price-sensitive developers.
The OpenRouter data tells the story. In the week of February 9-15, 2026, MiniMax M2.5 alone generated 3 trillion tokens of global API traffic, representing approximately 40% of all tokens routed through OpenRouter that week. By March, M2.5 and its successor variants had helped push total Chinese model token consumption to 9.22 trillion per week, nearly double the American total of 4.93 trillion.
| Model | Input Price ($/M tokens) | Output Price ($/M tokens) | Coding Benchmark | OpenRouter Share (Feb 2026) |
|---|---|---|---|---|
| GPT-4 Turbo | $10.00 | $30.00 | 92.3% | 8.2% |
| Claude Opus 4.6 | $5.00 | $15.00 | 94.1% | 12.5% |
| MiniMax M2.5 | $0.30 | $0.90 | 90.3% | 40.1% |
| DeepSeek V3.2 | $0.27 | $0.80 | 88.7% | 15.3% |
| GLM-5 (Zhipu) | $0.50 | $1.50 | 86.2% | 8.9% |
| Kimi K2.5 | $0.60 | $1.80 | 89.5% | 10.2% |
*Table 2: API pricing and performance comparison for top-tier models on OpenRouter, February 2026. Coding benchmark = HumanEval pass rate. OpenRouter share = weekly token consumption percentage. Data: OpenRouter, Artificial Analysis, company disclosures.*
Factor 2: Enterprise MaaS Platform Acceleration
MiniMax's B2B platform revenue grew 197.8% in 2025 and continued accelerating into 2026. The company positioned itself as a Model-as-a-Service (MaaS) provider for Southeast Asian, Middle Eastern, and Latin American developers who needed GPT-4-class performance at 1/10th the price.
By May 2026, the enterprise client count had crossed 1 million—up from approximately 200,000 just six months prior. This was not primarily Chinese domestic growth. OpenRouter's user demographics show that 47.17% of its users are American developers, while Chinese developers represent only 6.01%. MiniMax's API growth was overwhelmingly driven by international adoption.
| Revenue Segment | 2024 | 2025 | 2025 Share | Growth Driver |
|---|---|---|---|---|
| AI Native Products (C端) | $21.8M | $53.1M | 67.1% | Talkie, Hailuo AI, MiniMax Voice |
| — Talkie/星野 | $19.5M | $18.8M | 23.7% | User growth +62%, ARPPU challenges |
| — Hailuo AI | $2.3M | $17.5M | 22.1% | Video AI boom, subscription model |
| — MiniMax Voice | — | $1.1M | 1.3% | Speech synthesis, 12% paid rate |
| Open Platform (B端) | $8.7M | $26.0M | 32.9% | API calls, MaaS, enterprise |
| Total Revenue | $30.5M | $79.0M | 100% | +159% YoY |
*Table 3: MiniMax revenue breakdown by segment. The Hailuo AI video tool emerged as the fastest-growing product, while Talkie's revenue growth slowed despite user expansion. Data: company prospectus, Guosen Securities research.*
The User Engine: From 200 Million to 300 Million
Talkie: The AI Companion That Conquered America
Talkie, MiniMax's AI companion app (branded as 星野 domestically), has been the company's most visible consumer success. By May 2026, the product had accumulated 1.47 billion total user interactions and maintained approximately 20 million monthly active users—comparable to Character.AI, its primary American competitor.
The product's success is rooted in a specific design choice: card-based character collection. Users don't just chat with AI characters—they collect them, level them up, and trade rare cards. This gamification layer, borrowed from Japanese gacha mechanics and Chinese web novel culture, created retention metrics that pure chatbot apps cannot match.
Talkie's user base is heavily American: 55.18% of daily active users are from the United States, with the UK, Canada, Australia, and the Philippines as secondary markets. The product's revenue model is primarily advertising-supported with a subscription tier, though the company has faced challenges converting the large user base into high ARPPU (Average Revenue Per Paying User).
| Product | Total Users | MAU | DAU (est.) | Top Market | Revenue Model |
|---|---|---|---|---|---|
| Talkie | 147M+ | 20M | ~3.5M | USA (55%) | Ads + subscriptions |
| Hailuo AI | 50M+ | 5.6M | ~900K | China + SEA | Subscription tiers |
| MiniMax Voice | 3.7M | 494K | ~80K | Global | Pay-per-use + plans |
| Open Platform | — | — | — | Global (API) | Token-based billing |
| Combined | 300M+ | 27M+ | ~4.5M | International | Mixed |
*Table 4: MiniMax product portfolio user metrics, May 2026. Data: company disclosures, Sensor Tower, industry estimates.*
Hailuo AI: The Video Generation Dark Horse
While Talkie gets the headlines, Hailuo AI (海螺AI) has become MiniMax's revenue engine. The video generation tool, positioned as a competitor to Kling, Runway, and Pika, generated $17.5 million in revenue in the first nine months of 2025—a product that had barely existed in 2024.
Hailuo's success is tied to two trends: the AI short drama explosion (China's AI-generated short drama market grew from $100M to $650M in Q1 2026) and the global demand for low-cost video generation. At approximately $0.03 per second of generated video, Hailuo undercut Runway Gen-3 by roughly 70% while achieving comparable quality on AAVA benchmarks.
The product's user base of 5.6 million monthly active users is smaller than Talkie's, but its monetization is more efficient: subscription tiers range from $9.99 to $99.99 monthly, with a 12% paid conversion rate among active users.
The Financial Reality: Growth vs. Profitability
The $250 Million Loss on $79 Million Revenue
MiniMax's financials reveal the brutal economics of AI model development. In 2025, the company spent approximately $250 million—more than triple its revenue—on model training, compute infrastructure, and global expansion.
The cost structure is dominated by GPU cluster rentals and training runs. MiniMax operates its own inference infrastructure across data centers in China, Singapore, and the United States. The company's research team of approximately 800 engineers (implied by hiring data and competitor benchmarks) represents a significant fixed cost.
| Cost Category | 2025 ($M) | % of Revenue | Trend |
|---|---|---|---|
| Compute (training + inference) | ~$120 | 152% | Decreasing per-token |
| R&D personnel | ~$80 | 101% | Growing with team size |
| Sales & marketing | ~$25 | 32% | High for user acquisition |
| G&A | ~$15 | 19% | Stable |
| Content/ licensing | ~$10 | 13% | Legal settlements rising |
| Total Costs | ~$250 | 316% | Improving margin |
| Net Loss | ~$250 | 316% | Narrowing |
*Table 5: MiniMax estimated cost structure for 2025. Exact figures not fully disclosed; derived from prospectus data, comparable company filings, and industry analysis. The compute-to-revenue ratio of 152% is typical for pre-profitability AI infrastructure companies.*
The gross margin improvement from 12.2% to 25.4% suggests that the company is achieving operational leverage: as token volume grows, the marginal cost per token declines. However, the path to net profitability requires either a dramatic reduction in training costs or a significant revenue acceleration that outpaces infrastructure investment.
The IPO Timing: Why Now?
MiniMax filed for A-share IPO on the Shanghai Stock Exchange's STAR Market (科创板) on June 1, 2026. The timing is strategic: the company needs capital to fund the next generation of model training (M3 and beyond) while its growth metrics are at their most impressive.
China's STAR Market has become the preferred listing venue for high-tech companies, with more relaxed profitability requirements than the main board. The market has already hosted MiniMax's competitor Zhipu AI (listed January 2026) and is preparing for additional AI listings. MiniMax's Hong Kong listing in January 2026 (at HK$165 per share, now trading at HK$821) demonstrated that public market investors would pay a premium for AI growth stories.
| Company | Listing Date | IPO Price | Current Price | Market Cap | Revenue | P/S Ratio |
|---|---|---|---|---|---|---|
| MiniMax (HK) | Jan 2026 | HK$165 | HK$821 | HK$257.5B (~$33B) | $79M | ~418x |
| Zhipu AI (HK) | Jan 2026 | HK$142 | HK$680 | HK$180B (~$23B) | $45M | ~511x |
| Character.AI (US, private) | — | — | — | ~$10B (est.) | ~$80M | ~125x |
| Anthropic (US, private) | — | — | — | ~$61.5B | ~$2B | ~31x |
| OpenAI (private) | — | — | — | ~$300B | ~$5.5B | ~55x |
*Table 6: AI company valuation comparison. MiniMax and Zhipu's Hong Kong valuations reflect the premium Chinese public markets place on AI growth. P/S ratios are extraordinarily high by traditional standards. Data: exchange filings, Bloomberg, industry estimates.*
The M3 Model: Technical Foundation for the Next Phase
Launched the Same Day as the IPO Filing
MiniMax M3, released on June 1, 2026, represents the company's attempt to move upmarket. The model is positioned as a frontier-level system for coding and agentic tasks—the two highest-value AI application categories.
The timing of the M3 launch alongside the IPO filing is not coincidental. It signals to investors that MiniMax is not merely a low-cost API provider but a genuine frontier lab capable of competing with OpenAI, Anthropic, and Google on the most demanding tasks.
Technical specifications disclosed by the company include:
- Mixture-of-Experts (MoE) architecture with 1.2 trillion total parameters, 32 billion active per forward pass
- 2 million token context window for long-document processing
- Multi-modal input: text, image, audio, and video understanding in a single model
- Agent capabilities: tool use, multi-step planning, and code execution with sandboxed environments
| Model | Parameters (total) | Context Window | Coding Score | Agent Benchmark | Price ($/M input) |
|---|---|---|---|---|---|
| GPT-5.5 | Unknown | 2M | 94.5% | 87.2% | $10.00 |
| Claude Opus 4.8 | Unknown | 2M | 93.8% | 89.1% | $5.00 |
| Gemini 3.5 Pro | Unknown | 1M | 92.1% | 85.3% | $3.50 |
| MiniMax M3 | 1.2T (MoE) | 2M | 91.4% | 84.7% | $1.20 |
| DeepSeek V4 Pro | 671B (MoE) | 1M | 90.8% | 82.1% | $0.50 |
| Kimi K2.6 | Unknown | 2M | 89.2% | 80.5% | $1.80 |
*Table 7: Frontier model comparison, June 2026. Coding score = HumanEval + MBPP composite. Agent benchmark = SWE-bench + GAIA composite. Prices are API input rates. Data: company disclosures, benchmark leaderboards, AI in China editorial analysis.*
MiniMax M3's pricing at $1.20 per million input tokens—4x cheaper than GPT-5.5, 60% cheaper than Claude Opus—reflects the company's continued commitment to the "density law" strategy: achieving near-frontier performance at a fraction of the cost. The question is whether this pricing can generate sufficient margin to fund the next training run, estimated at $100-200 million for a model of this scale.
Social Comments: What Chinese Netizens Are Saying
"MiniMax的估值已经比Anthropic的一半还高了，但收入只有Anthropic的4%。这泡沫比2021年的互联网还大。" ("MiniMax's valuation is already half of Anthropic's, but its revenue is only 4% of Anthropic's. This bubble is bigger than the 2021 internet boom.")
— 雪球 user @科技投资老兵, 12,000 upvotes
"Talkie确实好玩，我花了两百多买卡牌。但问题是，这种AI陪伴产品能不能持续让用户付费？新鲜感过了怎么办？" ("Talkie is genuinely fun. I spent over ¥200 on cards. But the question is whether AI companion products can sustain user payments. What happens when the novelty wears off?")
— Bilibili user @虚拟世界漫游者, 8,500 likes
"Hailuo做视频确实便宜，我做自媒体一个月省了三千块外包费。但生成的人物手指还是经常出问题。" ("Hailuo is genuinely cheap for video. I save ¥3,000 a month on outsourcing as a content creator. But generated characters still have finger problems often.")
— Xiaohongshu user @短视频创作者小鹿, 15,000 likes
"A股IPO就是去割韭菜。250亿美元亏损还上市，科创板的韭菜真好割。" ("A-share IPO is just harvesting retail investors. Listing with $250 million in losses—STAR Market retail investors are easy prey.")
— Weibo user @财经冷眼, 22,000 retweets
"MiniMax M3的代码能力我测试了，写Python确实比Kimi强，但比Claude还是差一截。1.2美元的价格倒是真香。" ("I tested MiniMax M3's coding ability. It's better than Kimi for Python, but still a step below Claude. The $1.20 price is genuinely attractive.")
— Zhihu user @全栈工程师老张, 6,000 upvotes
"MiniMax出海的成功证明了一件事：中国AI不是靠抄袭，是靠性价比。老外也不是傻子，便宜好用就是王道。" ("MiniMax's overseas success proves one thing: Chinese AI doesn't win by copying, it wins by value. Foreigners aren't stupid—cheap and good is king.")
— WeChat Channels user @出海观察家, 9,200 likes
Competitive Landscape: China's AI Export Race
The "Six Little Tigers" and Beyond
MiniMax is one of the "AI Six Little Tigers" (AI六小虎)—the six most prominent Chinese AI startups: MiniMax, Zhipu AI, Moonshot AI (Kimi), Baichuan AI, 01.AI, and StepFun. Each has pursued a different international strategy, with MiniMax clearly leading in user scale and revenue diversification.
| Company | Primary Model | C端 Product | B端 Platform | International Revenue | IPO Status |
|---|---|---|---|---|---|
| MiniMax | M3 (MoE) | Talkie, Hailuo | Open Platform | 73% | HK listed, A-share filing |
| Zhipu AI | GLM-5 | Qingyan APP | Zhipu Platform | ~35% | HK listed |
| Moonshot (Kimi) | Kimi K2.6 | Kimi APP | Kimi API | ~15% | Private, rumored HK 2026 |
| Baichuan | Baichuan 4 | — | Enterprise API | ~10% | Private |
| 01.AI | Yi-Large | Wanzhi | Yi Platform | ~20% | Private |
| StepFun | Step 3.5 | Yuewen | Step Platform | ~12% | Private |
| DeepSeek | V4 Pro | — | DeepSeek API | ~25% | No IPO plans |
*Table 8: China's "AI Six Little Tigers" competitive comparison. International revenue share varies dramatically, with MiniMax as the clear leader in global monetization. Data: company disclosures, analyst reports, industry estimates.*
The ByteDance Shadow
No analysis of Chinese AI would be complete without acknowledging ByteDance. Doubao (豆包), ByteDance's AI assistant, reached 449 million monthly active users by March 2026—more than 15x MiniMax's combined MAU. Volcano Engine, ByteDance's enterprise cloud, holds 49.5% of China's MaaS market.
However, ByteDance's international AI strategy is constrained by geopolitical factors. TikTok's regulatory challenges in the United States have made the company cautious about promoting Doubao or other AI products aggressively in Western markets. This creates a strategic window for MiniMax: it can be the "Chinese AI company that isn't ByteDance"—familiar enough to benefit from China's AI reputation, but small enough to avoid direct regulatory targeting.
| Company | Domestic MAU | International MAU | Enterprise MaaS Share | Primary International Risk |
|---|---|---|---|---|
| ByteDance (Doubao) | 449M | Minimal | 49.5% | TikTok regulatory spillover |
| Alibaba (Tongyi) | 120M | 30M | 28.0% | Cloud competition, US sanctions |
| MiniMax | 8M | 19M | ~5% | Copyright litigation, data privacy |
| Baidu (Wenxin) | 85M | 15M | 8.2% | Legacy reputation, model quality |
| Tencent (Hunyuan) | 60M | 10M | 6.7% | Conservative AI investment |
*Table 9: Chinese AI company international presence comparison. MiniMax punches above its weight in international MAU relative to its domestic footprint. Data: QuestMobile, IDC, company disclosures.*
What's Next: The Path to Profitability
Three Scenarios for 2026-2027
MiniMax's IPO filing requires a credible profitability narrative. Based on current trajectory and comparable company paths, we model three scenarios:
| Scenario | Revenue 2026E | Revenue 2027E | Gross Margin 2027E | Net Profit 2027E | Probability |
|---|---|---|---|---|---|
| Bull: Hailuo + API hypergrowth | $300M | $600M | 45% | +$50M | 25% |
| Base: Steady growth, cost discipline | $200M | $350M | 38% | -$80M | 50% |
| Bear: User churn, price war | $150M | $220M | 30% | -$150M | 25% |
*Table 10: MiniMax profitability scenarios. The bull case requires Hailuo to become the dominant global video generation tool and the API platform to maintain 100%+ growth. The bear case reflects a potential price war with DeepSeek and open-source alternatives. Data: AI in China editorial modeling.*
The Key Risks
1. Copyright litigation: MiniMax faces a $75 million lawsuit from Hollywood studios alleging unauthorized use of copyrighted video in training data. A negative ruling could force model retraining and damage the Hailuo brand.
2. Data privacy regulation: Talkie's collection of intimate user conversations has attracted scrutiny from EU and US regulators. A GDPR or COPPA enforcement action could restrict the app's largest market.
3. Geopolitical escalation: US tariffs on AI-related services (currently 25% on chip imports) could expand to API services, directly impacting MiniMax's pricing advantage.
4. Model commoditization: As open-source models (Llama 4, Qwen 3.6, DeepSeek) approach frontier performance, the premium API market may compress, forcing further price cuts.
Related Articles
- China's AI Models Dominate Global API Traffic: The Token Export Revolution
- DeepSeek's $51.5B Valuation: The Anti-Startup Startup Bows to Capital
- China's AI Drama Revolution: How ByteDance Built a $650M Short-Content Empire
- Alibaba's Zhenwu M890 AI Chip: China's Silicon Sovereignty Push
*Published June 2, 2026. Data current as of May 31, 2026. Financial figures based on MiniMax prospectus filings, Bloomberg interviews, and Guosen Securities research. Revenue and loss estimates are editorial projections unless sourced from official company disclosures. Valuation figures reflect Hong Kong stock exchange closing prices.*
Editor at AI in China. Tracking Chinese AI companies, funding rounds, and the technologies reshaping global tech. More about me.