explainx / blog
DeepSeek, Zhipu, Moonshot, MiniMax, Qwen, and the Six Tigers โ ranked by models, funding, open weights, and enterprise fit. Updated June 2026 with pricing and how to try each stack.
Jun 29, 2026
On June 28, 2026, a thread with 200K+ views argued China's AI playbook is simple: ship great models for free, export cheap inference powered by low-cost electricity, and wait for Huawei to close the chip gap. Here's what the skeptics push back on โ and what it means if they're wrong.
Jun 29, 2026
America outspends China 23-to-1 on private AI investment but barely leads on model benchmarks. US and Chinese AI startups are running different races โ this guide maps funding, strategy, moats, and where each side actually wins.
Jun 28, 2026
Chinese tech firms and analysts warned in late June 2026 that the AI bubble is close to bursting โ pointing to inflated GPU valuations, weak enterprise ROI, and a widening gap between AI infrastructure spending and actual revenue. Here's the full breakdown.
China's AI industry in mid-2026 is not one company or one model โ it is a layered ecosystem of research labs, venture-backed startups, and big-tech cloud platforms that together account for roughly 45% of OpenRouter token traffic, up from under 2% a year earlier. The Stanford HAI AI Index 2026 reports the USโChina top-model performance gap has collapsed to 2.7% while China spent roughly 23ร less on private AI investment ($12.4B vs $285.9B in 2025).
That efficiency story is why developers, procurement teams, and founders keep asking the same question: who actually matters in Chinese AI, and which lab fits which job?
This guide answers that โ company by company, with models, positioning, pricing anchors, and honest limits. It complements explainx.ai's coverage of the China AI playbook and the US vs Chinese startup comparison.
| If you needโฆ | Start here | Flagship model (Jun 2026) | Why |
|---|---|---|---|
| Cheapest frontier coding API | MiniMax, StepFun | M2.7, Step 3.5 Flash | ~$0.10โ0.30 input; strong SWE-bench scores |
| Open weights + self-host | DeepSeek, Zhipu | DeepSeek V4 Pro, GLM-5.2 | Commercial-friendly licenses; Hugging Face weights |
| Long-context agents | Moonshot | Kimi K2.7 | Agent Swarm; powers Cursor Composer integrations |
| Broadest model family | Alibaba (Qwen) | Qwen 3.7-Max, Qwen 3.6 Plus | 900M+ HF downloads; 1M context tiers |
| Enterprise MaaS in China | Zhipu (Z.ai) | GLM-5.2 | 9/10 top Chinese internet firms; Huawei Ascend stack |
| Consumer AI apps | ByteDance, MiniMax | Doubao (Seed), Talkie | 155M+ WAU on Doubao; MiniMax 70% intl revenue |
| Domestic chip independence | Zhipu, DeepSeek, Baidu | GLM-5, ERNIE 5.0 | Trained/served on Ascend or Kunlun silicon |
| Highest OpenRouter volume | Xiaomi | MiMo-V2-Pro | ~21% platform share; free preview drove IDE adoption |
Chinese AI splits into three tiers that behave differently in procurement and developer adoption:
Tier 1 โ Big Tech Labs (cloud + consumer scale)
Alibaba (Qwen) ยท ByteDance (Doubao/Seed) ยท Baidu (ERNIE) ยท Tencent (Hunyuan)
Tier 2 โ Independent Startups ("Six Tigers" + DeepSeek)
Zhipu ยท Moonshot ยท MiniMax ยท Baichuan ยท StepFun ยท 01.AI ยท DeepSeek
Tier 3 โ Hardware-adjacent & niche
Xiaomi (MiMo) ยท Huawei (Ascend stack) ยท iFlytek (Spark) ยท Kuaishou (KwaiKAT)
Six Tigers (ๅ ญๅฐ่) is the label Chinese media and investors use for six leading independent startups. DeepSeek is almost always discussed in the same breath but sits slightly apart โ a Hangzhou research lab that prioritizes open-weight efficiency over consumer apps.
Four Dragons (ๅๅฐ้พ) is a newer investor label for the four highest-valued independents: DeepSeek, Zhipu, MiniMax, and Moonshot โ combined valuations reportedly exceeded 1 trillion yuan (~$140B) by early 2026 per China Biz Insider reporting.
These are the companies most Western developers mean when they say "Chinese AI startups."
| Founded | 2023, Hangzhou |
| Focus | Research-first; open weights; training efficiency |
| Key models | DeepSeek V4 Pro, V4-Flash, R1 reasoning line |
| Differentiator | MoE architecture, sparse attention (DSA), reported ~$6M-class training economics |
| Try it | API docs ยท Hugging Face collection |
DeepSeek triggered the January 2026 market shock by demonstrating frontier-class models at a fraction of Western training cost. Its strategy aligns with the commoditization playbook: release weights, undercut API pricing, force global repricing.
Best for: Self-hosters, researchers, cost-sensitive agent pipelines, teams building on DeepSeek TUI agents.
Limits: Compliance documentation thin vs US cloud providers; geopolitical scrutiny on any China-hosted API route.
| Founded | 2019, Beijing (Tsinghua University spinout) |
| Focus | Enterprise MaaS, developer API, government-linked deployments |
| Key models | GLM-5.2, GLM-5, ChatGLM |
| Differentiator | MIT-licensed weights; Huawei Ascend end-to-end; 9/10 top Chinese internet firms as customers |
| Try it | Z.ai ยท How to run GLM on Ollama |
Zhipu is the most commercially mature of the independents. CEO Zhang Peng explicitly positions the company as "China's Anthropic" โ base-model moat, API-first revenue, developer ecosystem flywheel. GLM-5.2 matched Claude Mythos on security benchmarks in late June 2026.
Reported scale (per 36Kr coverage of Zhipu's 2026 outlook): ~2.5B yuan ARR, 240,000+ paying developers, GLM models deployed on Google Vertex AI and AWS Bedrock.
Best for: Enterprise API buyers, self-hosters who want MIT license, security-adjacent evals, agent harness integration.
Limits: ~76ร smaller ARR than Anthropic โ ecosystem depth still catching up; US regulated buyers face vendor review.
| Founded | 2023, Beijing (Tsinghua; Yang Zhilin, CMU alum) |
| Focus | Kimi chatbot, long-context reasoning, agent orchestration |
| Key models | Kimi K2.7, Kimi K2.5 |
| Differentiator | Agent Swarm (100-agent coordination); 262Kโ1M context; Cursor Composer 2 base model |
| Try it | kimi.com ยท OpenRouter ยท open weights on Hugging Face |
Moonshot built China's second-most-popular consumer chatbot after ByteDance's Doubao. Its K2 line targets agentic coding and long-document workflows โ students, lawyers, and white-collar workers were early power users for 200K+ character inputs.
Growth has been explosive: reporting cited $100M ARR in March 2026 and $200M+ in April 2026, with month-over-month doubling after K2.5 launch. Alibaba is a major investor.
Best for: Long-context RAG, multi-agent orchestration, coding agents, teams evaluating Kimi vs Claude on planning tasks.
Limits: Some reporting suggests Moonshot scaled back pure foundation research in 2025โ2026 as key talent moved; verify current model roadmap before multi-year bets.
| Founded | 2021, Shanghai |
| Focus | Consumer apps (Talkie/Xingye) + enterprise M2 coding models |
| Key models | MiniMax M2.7, M2.5, Hailuo video |
| Differentiator | ~70% revenue from international markets; self-evolving agent architectures |
| Try it | minimax.io ยท OpenRouter |
MiniMax is the consumer-facing tiger. Talkie (Glow internationally) is a character-AI chatbot with strong overseas traction. On the developer side, the M-series shipped four major releases in six months, with M2.7 hitting ~56% SWE-Pro at roughly 50ร lower cost than Claude Opus on comparable agent workloads.
Listed on Hong Kong Stock Exchange alongside Zhipu โ a signal Beijing wants AI startups to access public capital.
Best for: Agentic coding at low cost, consumer AI prototypes, international app builders.
Limits: Flagship weights often closed; compliance docs weaker than Alibaba Cloud or Tencent Cloud.
| Founded | 2023, Shanghai |
| Focus | Fast MoE inference, Step series language models |
| Key models | Step 3.5 Flash, Step-2 (1T params) |
| Differentiator | 196B MoE with 11B active; 350 tok/s on Hopper; top free-tier OpenRouter volume |
| Try it | OpenRouter ยท StepFun platform |
StepFun is the under-discussed performer in Western coverage. Step 3.5 Flash sits at ~$0.10/$0.30 per million tokens with a strong free preview โ #3 on OpenRouter's free-model leaderboard in Q2 2026.
Talent pedigree includes researchers from Google, Microsoft, and Chinese big tech. Step-2's trillion-parameter scale ranked competitively on LiveBench against DeepSeek and OpenAI entries.
Best for: Cost-sensitive batch inference, prototyping, teams that want Hopper-trained models (not Ascend-only).
| Founded | March 2023, Beijing |
| Focus | Open-weight LLMs for enterprise and verticals |
| Key models | Baichuan 4, Baichuan-M2 |
| Differentiator | Microsoft SVP-level leadership; strong Chinese NLP |
| Try it | Hugging Face ยท Baichuan API |
Baichuan was an early Six Tigers member with talent from Microsoft, Huawei, Baidu, and Tencent. Valuation reportedly reached ~$2.8B in 2024 funding rounds.
Market share on global routers is smaller than Qwen or GLM, but Baichuan remains relevant for domestic enterprise buyers who want an alternative to Alibaba's ecosystem.
Best for: Chinese-language enterprise NLP, domestic compliance paths.
| Founded | 2023 (Sinovation Ventures) |
| Focus | Yi model family, enterprise AI platform |
| Key models | Yi-Lightning, Yi-Large |
| Differentiator | Kai-Fu Lee founder brand; global developer outreach |
| Try it | 01.ai ยท Hugging Face Yi collection |
01.AI launched with significant fanfare as Lee's post-Sinovation AI venture. The Yi models competed on open-weight leaderboards in 2024โ2025. By 2026, volume share has consolidated toward Qwen, DeepSeek, and GLM โ but 01.AI remains in the Six Tigers conversation for brand and enterprise partnerships.
Best for: Teams already on Yi fine-tunes; Kai-Fu Lee ecosystem connections in Asia.
These companies have distribution, cloud infrastructure, and consumer apps that startups cannot match.
The scale leader. Qwen holds 942M+ Hugging Face downloads as of early 2026 โ more than the next eight competitors combined per industry reporting.
| Model tier | Role |
|---|---|
| Qwen 3.7-Max | Long-horizon agent frontier |
| Qwen 3.6 Plus | 1M context flagship; #2 OpenRouter coding share |
| Qwen 3.5 Flash | Ultra-cheap batch ($0.065/$0.26 per M tokens) |
| Qwen 3 Coder Next | IDE-optimized coding variant |
Alibaba Cloud International gives Qwen the least painful cross-border enterprise path among Chinese providers โ EU regions, established cloud certs.
Investor angle: Alibaba leads rounds in Moonshot and holds stakes across the ecosystem โ the SoftBank-of-Chinese-AI pattern.
ByteDance runs China's most-used AI consumer app โ Doubao with 155M+ weekly active users. The Seed model family powers Doubao with pricing reportedly ~10ร below GPT-5.2 on comparable tasks.
Seed 2.0 Lite and Mini extend into ultra-cheap tiers. Volcano Engine (็ซๅฑฑๅผๆ) is the enterprise cloud surface.
Best for: Consumer AI at Chinese scale, short-video + AI integration (TikTok/Douyin ecosystem).
Baidu pivoted aggressively to open source in 2025 โ from zero Hugging Face releases in 2024 to 100+ in 2025. ERNIE 5.0 is a 2.4T-parameter omnimodal flagship trained on Baidu Kunlun chips and integrated with China's dominant search engine.
Best for: Search-native AI, domestic cloud buyers already on Baidu AI Cloud.
Tencent keeps Hunyuan primarily inside its ecosystem โ WeChat, Tencent Cloud, gaming, and enterprise SaaS. Minimal OpenRouter footprint by design; maximum leverage across Tencent's 1B+ WeChat users.
Best for: Teams already on Tencent Cloud; WeChat mini-program AI.
A smartphone company holds ~21% of OpenRouter weekly tokens โ 3ร OpenAI's share โ because MiMo-V2-Pro shipped with free preview, 1M+ context, and IDE partnerships. Three MiMo variants (Pro, Omni, Flash) cover flagship, multimodal, and ultra-cheap tiers.
Lesson: Distribution beats benchmark rank. MiMo ranks #10 on intelligence indexes but #1 by usage.
Huawei does not ship consumer LLMs at the same tier as Qwen โ but Ascend chips are the domestic silicon story. GLM-5 and upcoming DeepSeek V4 train on Ascend. For Chinese state and SOE buyers, hardware independence beats a benchmark point.
Dominates speech, education, and public-sector AI in China. Limited global developer relevance unless you need Chinese voice + language bundled.
| Company | Valuation / scale signal | Revenue model | Open weights? | Best workload |
|---|---|---|---|---|
| DeepSeek | ~$45B reported | API + open weights | Yes (strong) | Research, self-host, cost disruption |
| Zhipu | HK-listed; ~377B yuan mkt cap reported | Enterprise MaaS | Yes (MIT) | B2B API, security, Ascend deploys |
| Moonshot | ~$20B+ reported | Kimi consumer + API | Partial (K2 line) | Long context, agents, coding |
| MiniMax | HK-listed; ~223B yuan reported | Consumer + API | Partial | Agentic coding, Talkie apps |
| StepFun | Private | API | Partial | Cheap/fast MoE inference |
| Baichuan | ~$2.8B (2024 round) | Enterprise | Yes | Chinese enterprise NLP |
| 01.AI | Private (Lee-backed) | API + Yi weights | Yes | Yi ecosystem, Asia enterprise |
| Alibaba/Qwen | Public parent | Cloud + API | Yes (broad) | Scale, downloads, cloud certs |
| ByteDance | Private giant | Consumer + Volcano | Mixed | Doubao consumer scale |
| Xiaomi/MiMo | Public parent | API | Mixed | IDE volume, cheap coding |
Valuation figures from public reporting and investor materials; treat as directional, not audited.
No โ officially the Six Tigers are Zhipu, Moonshot, MiniMax, Baichuan, StepFun, and 01.AI. DeepSeek is a separate Hangzhou lab that outgrew the label on global impact. In practice, every "Chinese AI startups" list includes DeepSeek first.
For Claude Code replacements during the Fable export ban, teams benchmarked GLM-5.2, Kimi K2.7, and OpenRouter Fusion. Cursor integrated Kimi K2.5 as Composer 2's base. There is no universal winner โ run your task suite.
Self-hosted open weights: manageable risk with standard security review.
Chinese-hosted APIs with sensitive data: hard no for most regulated US/EU buyers. NIST and USCC reports flag higher cyber-risk scores and content-policy filtering on some Chinese models vs US counterparts.
Pragmatic middle path: Route non-sensitive coding and batch work to Chinese APIs; keep customer PII on US frontier stacks. See enterprise open-source alternatives guide.
Use this checklist before production routing:
# Quick eval pattern โ swap model ID per vendor
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/deepseek-v3.2",
"messages": [{"role": "user", "content": "Your real production prompt here"}]
}'
Run the same prompt across GLM, Kimi, Qwen, and your US baseline. Compare latency, cost, and output quality โ not marketing tables.
Chinese AI in 2026 is a mature, consolidated market โ not a single DeepSeek surprise. Ten providers matter for production; six startups carry the independent innovation label; four big-tech clouds own distribution.
The right choice is almost never "pick one Chinese model." It is tier by workload: Qwen or MiMo for cheap coding volume, GLM or DeepSeek for self-hosted enterprise, Kimi for long-context agents, MiniMax for agentic price-performance, Alibaba or Tencent cloud if you need compliance documentation.
The performance gap with the US is measured in percentage points. The price gap is measured in multiples. That is why 45% of OpenRouter traffic already routes through Chinese providers โ and why every serious AI stack in 2026 needs a China-aware routing layer.
Company valuations, ARR figures, and OpenRouter share data reflect public reporting and platform statistics as of June 29, 2026. Verify live pricing and terms on each vendor's site before procurement decisions.