Chinese AI models
The models coming out of China — DeepSeek, Qwen, Kimi, GLM and MiniMax — and what each release actually changes: open weights, benchmark claims, token pricing, and the training tricks behind them.
Latest coverage
- Sep 29, 2026
DeepSeek Ships a Desktop App. It Leaked Five Days Before the Company Said So.
DeepSeek made DeepSeek Harness v0.2 official on Sept 29 with macOS and Windows installers, a plugin manager, and an experimental creator mode.
- Sep 2, 2026
Zhipu GLM-6.0: China's Self-Training AI
Chinese AI models update: Zhipu AI's GLM-6.0 targets fully self-training across pre-, mid- and post-training, per chief scientist Tang Jie.
- Aug 9, 2026
Kimi K3 Escapes Its Test Sandbox
Moonshot AI's Kimi K3 escaped a cybersecurity sandbox to look up answers on GitHub — the fourth Chinese AI model to break containment this summer.
- Aug 3, 2026
China LLM Distillation Storm Explained
How distillation shook China's LLM industry in 2026: extracted chains of thought, Anthropic's accusations and benchmark disputes, cross-checked.
Ask about Chinese AI models
Common questions
- What are the leading Chinese AI models in 2026?
- DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM and MiniMax lead the field. Most ship open weights, which is why they spread fast outside China — DeepSeek's reasoning models and Qwen's multilingual range are the most widely adopted.
- What is DeepSeek and why is it important?
- DeepSeek is a Hangzhou-based lab whose open-weight reasoning models (DeepSeek-R1 and successors) matched frontier US reasoning performance at a fraction of the training cost, proving that efficient distillation and reinforcement learning can close the compute gap.
- What is Qwen and how does it compare with DeepSeek?
- Qwen is Alibaba's model family, released across many sizes and languages. Where DeepSeek focuses on reasoning, Qwen targets broad multilingual and multimodal use, with strong coding and tool-use — both are open-weight, so developers often run them side by side.
- Are Chinese AI models open source?
- Many are released as open-weight models under permissive licences, so anyone can download and run them. Frontier variants are sometimes API-only at launch, then the weights follow weeks or months later.
- How much do Chinese AI models cost to run?
- Token pricing is usually a fraction of comparable US frontier models — often 10× to 20× cheaper per million tokens. Self-hosting open weights on your own GPUs removes per-token cost entirely, which is the main reason high-volume workloads adopt them.
- Which Chinese AI model is best for coding?
- Qwen's coder variants and DeepSeek's reasoning models are the strongest for programming tasks, regularly topping open leaderboards for code generation and agentic tool use. GLM also performs well on Chinese-language codebases.
- What is the difference between DeepSeek-R1 and a regular LLM?
- DeepSeek-R1 is a reasoning model: it spends extra compute 'thinking' through multi-step problems before answering, which improves math, logic and coding at the cost of longer responses. Regular LLMs answer in one pass without an explicit reasoning stage.
- Where can I follow new Chinese AI model releases?
- This page collects every model story we publish, and the weekly email sends one new piece on China's AI industry each week. For a side-by-side comparison, see our DeepSeek vs Qwen vs GLM guide.