Open
Weights released by multiple leading labs
2024–2025
Home / Artificial Intelligence
Track 02 / 人工智能
The model is free. The strategy is anything but.
China's AI labs flipped the script: instead of guarding frontier models, they ship open weights and compete on efficiency and price. The shock isn't a single benchmark — it's the reset of what the world expects to pay for intelligence.
↓ Cost
Inference pricing pressure felt worldwide
MARKET TREND
Top-5
Chinese models in global open-model rankings
COMMUNITY BOARDS
Many
Domestic accelerators racing to fill the gap
CHINA KEY TRACK
The dynamics / 关键动力
Efficiency as a weapon.
When you can't always win on raw compute, you win on how cleverly you use it — and on who you let build on top.
01
Open weights as distribution
Releasing a capable model for free puts it inside every startup, university and enterprise overnight. Mindshare and ecosystem lock-in become the moat, not secrecy.
02
Architecture over brute force
Mixture-of-experts, smarter training pipelines and aggressive inference optimisation squeeze frontier-class results from constrained hardware budgets.
03
Applications at population scale
A billion-user domestic market is the ultimate testbed. Models harden fast against real demand across search, commerce, super-apps and devices.
Who to watch / 关注名单
The labs.
DeepSeek 深度求索
Qwen / Alibaba 通义千问
Zhipu AI 智谱
Moonshot 月之暗面
MiniMax 稀宇
Baidu ERNIE 文心
ByteDance 字节跳动
Questions answered / 问答
What people ask us about China AI.
What is DeepSeek and why did it matter?
DeepSeek is a Chinese AI lab whose open-weight reasoning models drew global attention for delivering strong performance at a fraction of the assumed training and inference cost — pressuring incumbent pricing and forcing a rethink of how expensive frontier AI has to be.
Are Chinese AI models actually open source?
Many leading families — DeepSeek, Alibaba's Qwen and others — release open weights you can download, run and fine-tune. "Open weights" isn't always full open-source, and licences differ, so the specific terms matter for commercial use.
Can China compete in AI despite chip restrictions?
So far, yes — by leaning into efficiency. Better architectures, training techniques and inference tricks let labs do more with constrained compute, while a domestic accelerator ecosystem matures alongside. It's a real constraint, not a hard ceiling.
Should Western developers use Chinese models?
Many already do, especially for cost-sensitive workloads and on-prem deployments where open weights shine. The considerations are licensing, data governance and support — which is exactly the kind of trade-off our briefings unpack.
Go deeper / 深入
Read the AI shift the way it actually happened.
We translate launches, benchmarks and licences into what they mean for builders and buyers — without the hype cycle.