# Chinese Frontier Labs
The five organisations the article treats as China's frontier, with the model names that appear in it.
| Lab | Model family | Notes |
|---|---|---|
| **DeepSeek** | V3, R1 | Spun out of the quant fund High-Flyer. Known for training efficiency on constrained hardware ([[DeepSeek Technical Differentiators]], [[DeepSeek Impact]]). |
| **Alibaba** | Qwen | Broadest open-weight family, many sizes and modalities; the most common base for derivative models on Hugging Face. |
| **Moonshot AI** | Kimi (K2, K2.5, K3) | Large MoE models with strong coding and agent performance; used as a base by US coding and legal AI products. |
| **ByteDance** | Doubao, Seed | Largest consumer user base; monetises largely through its commerce and content ecosystem. |
| **Zhipu AI** | GLM | Tsinghua-linked; open-weight GLM models used in coding agents. |
## Structural points
- **Competition inside China.** Five labs releasing [[Open Weights]] and leapfrogging each other every few months is a different dynamic from a small number of closed US labs. Much of the speed comes from China against China.
- **Open by default.** Releasing weights builds adoption where selling APIs abroad would hit procurement and trust barriers ([[Open Model Adoption Signals - OpenRouter and Hugging Face]]).
- **Monetisation differs.** Revenue is far below US frontier labs, with consumer products leaning on commerce and ads rather than subscriptions ([[Run Rate, ARR and Revenue Multiples]]).
- **Constraint-shaped engineering.** See [[Systems and Compiler Depth]] and [[AI Chip Export Controls]].
Related: [[Frontier Models]], [[China - Risky Bet or Unique Opportunity]]