Best Chinese AI models
Chinese labs now ship frontier-class models at a fraction of US flagship prices, and most of them publish open weights. These are the best Chinese AI models in 2026, ranked on real agentic-coding performance and value, from Moonshot, DeepSeek, Alibaba, MiniMax, Z.AI, and Xiaomi.
As of June 2026, picking the best Chinese model comes down to what you optimize for:
- Agentic coding: Kimi K2.6, the strongest Chinese model on CursorBench, with a free tier to try it
- Raw intelligence: Qwen3.7 Max, the highest Intelligence Index score in this list
- Value: MiniMax-M3 at $0.30 per million input tokens, or DeepSeek V4 for 1M context plus open weights
- High volume: DeepSeek V4 Flash or GLM 4.7 Flash, both around a tenth of a dollar or less per million input tokens
- Self-hosting: DeepSeek V4, Kimi K2.6, GLM 5.1, and Qwen3 Coder 480B all publish open weights
The best Chinese AI model right now is Kimi K2.6 from Moonshot AI: the strongest Chinese agentic coder at 47.6% on CursorBench 3.1, with open weights, a 262K context window, and even a free tier on OpenRouter.
For raw intelligence, Qwen3.7 Max scores highest on the Artificial Analysis Intelligence Index (56.6) of any Chinese model we track, and MiniMax-M3 (54.7) gets close at a remarkable $0.30 per million input tokens. DeepSeek V4 pairs a 1M token context window with open weights at $0.435.
The pattern across the board: Chinese labs match or approach frontier capability at 5 to 20 times lower prices, and unlike US flagships, most publish open weights you can self-host.
The frontier tier: models that compete directly with US flagships on capability, at much lower prices.
- 11Kimi K2.6Moonshot AIMoonshot's flagship: the strongest Chinese agentic coder (47.6% on CursorBench), with a free tier.Context262KInput$0.66/M
- 22Qwen3.7 MaxQwenAlibaba's frontier model: the highest Intelligence Index score of any Chinese model here.Context262KInput$0.8/M
- 33DeepSeek V4DeepSeekA 1M token context window and open weights at $0.435 per million input tokens.Context1.049MInput$0.435/M
- 44MiniMax-M3MiniMaxThe value flagship: near-frontier intelligence at $0.30 per million input tokens.Context205KInput$0.3/M
- 5Context203KInput$0.98/M
Open-weight specialists tuned for repositories and agentic edits.
- 11Qwen3 Coder 480B A35BQwenThe open-weight coding specialist: 1M context for big repositories, with a free tier.Context1.049MInput$0.22/M
Huge context windows and dependable quality at prices that make always-on agents and high-volume pipelines viable.
- 11DeepSeek V4 FlashDeepSeekThe cheapest 1M-context model anywhere: $0.098 per million input tokens.Context1.049MInput$0.09/M
- 2Context1.049MInput$0.105/M
- 3Context203KInput$0.06/M
Each model's Artificial Analysis Intelligence Index score against its blended price per 1M tokens. Toward the top right is more intelligence per dollar.
What is the best Chinese AI model in 2026?
Kimi K2.6 from Moonshot AI is the best Chinese model overall right now: the strongest Chinese agentic coder at 47.6% on CursorBench 3.1, with open weights and a free tier on OpenRouter. For raw intelligence, Qwen3.7 Max from Alibaba scores highest on the Artificial Analysis Intelligence Index among Chinese models we track.
Are Chinese models good for coding?
Yes, and coding agents are where they shine. Kimi K2.6 posts the best Chinese CursorBench result, and Qwen3 Coder 480B is an open-weight specialist with a 1M token context window for large repositories. Both are far cheaper than US flagships, which is why they dominate the top-model boards of OpenRouter-native agents like Hermes and OpenClaw.
Why are Chinese AI models so cheap?
A mix of efficient mixture-of-experts architectures, fierce domestic competition, and a strategy of publishing open weights so anyone can host them. Competition between hosting providers pushes prices down further. The result: models within reach of the frontier for $0.10 to $1 per million input tokens, versus $2.50 to $15 for US flagships.
Can you self-host Chinese AI models?
Most of them, yes. DeepSeek V4, Kimi K2.6, GLM 5.1, Qwen3 Coder 480B, and MiMo-V2.5 all publish open weights. The flagships need serious server hardware, but smaller open-weight options run on a workstation. The closed exceptions in this list are Qwen3.7 Max and MiniMax-M3, which are API only.
Models
9Updated
June 2026