China-US AI Gap Narrowing Rapidly as Chinese Models Capture Growing Share of Global Downloads
Nashnova编辑部
Chinese AI models are rapidly closing in on US leaders in coding and reasoning — while undercutting them on price and seizing the top spot in global downloads. That challenges the trillion-dollar valuations of OpenAI and Anthropic head-on.
How good is Kimi K3, really?
Moonshot AI's latest model, Kimi K3, scored 85.0% on a coding benchmark (Artificial Analysis Terminal-Bench 2.1), ranking sixth — all five spots above it belong to OpenAI and Anthropic.
This means → Chinese models have moved from "a tier behind" to "on the same leaderboard, within percentage points."
Independent evaluations show Kimi K3 approaches Anthropic's most advanced model in coding and reasoning, yet costs a fraction of the price.
In plain terms = near-equal capability at a far lower price — the combination US firms find hardest to answer.
Are global developers voting with their feet?
On Hugging Face, the world's largest model-download platform, Chinese models now account for the single largest share of generative-model downloads.
Alibaba's open-source Qwen series logged over 3 billion downloads in six months, surpassing Meta, Google, and Chinese peers to become the most-downloaded AI model family worldwide.
US companies including Airbnb, DoorDash, and Coinbase are already running Chinese models on their own servers.
This reflects a clear signal: once the performance gap narrows, price becomes the deciding factor — developers won't pay a brand premium.
Can the US block its way out of this?
Nearly 200 US companies have publicly opposed banning Chinese AI models, arguing the move would raise their own costs and weaken international competitiveness.
Kimi K3 and similar models are open-weight — their parameters are public, anyone can download, copy, and run them locally — making regulatory restrictions extremely hard to enforce.
In plain terms = software isn't chips. Chips can be stopped at customs; an open-source model, once uploaded, is available to everyone on the internet.
Can anyone stay ahead for long?
Ruslan Salakhutdinov, a Carnegie Mellon professor who supervised Moonshot AI founder Yang Zhilin's PhD, put it plainly: "They all get caught up to at some point."
Meta founder Mark Zuckerberg has called this AI race "one of the most intense competitions ever."
This means → in a fast-iterating AI field, there is no permanent technological moat — and the trillion-dollar valuations of OpenAI and Anthropic rest on an assumption of sustained leadership that is now being shaken.
Content is for reference only, not financial advice.