China's Moonshot and Alibaba unveil giant AI models they say rival GPT-5 and Claude
Two of China's biggest AI players released models this weekend that they claim match the best American systems, and plan to make them free to download. The gap between US and Chinese AI may be shrinking fast.
Key points
- Moonshot AI, a Beijing-based lab, released Kimi K3 on Friday, July 18, 2025, claiming it ranks above nearly every US model except OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5.
- Alibaba previewed Qwen3.8 over the weekend, calling it "second only to Fable 5" and describing it as one of the most powerful models available today.
- Moonshot says Kimi K3 has 2.8 trillion parameters, the internal numbers a model learns during training, making it the largest open-source AI system by that measure.
- Both Chinese companies plan to release their model weights, the files developers need to run and modify an AI model themselves, to the public for free.
- US export controls have blocked China from buying the most advanced chips, yet these models appear to approach the American frontier anyway.
Two of China's top AI companies dropped major announcements this weekend, and the timing was anything but accidental.
Moonshot AI, based in Beijing, unveiled a model called Kimi K3 on Friday. Alibaba, the Chinese e-commerce and technology giant, followed days later with a preview of Qwen3.8. Both companies say their models sit just behind the very best American systems, a claim that, if it holds up under independent testing, signals how fast the gap between US and Chinese AI is closing.
The comparison points they chose tell you everything. Moonshot says Kimi K3 trails only OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 (Anthropic's current flagship model) in its own tests. Alibaba calls Qwen3.8 "second only to Fable 5." Neither OpenAI nor Anthropic publish the technical specifications of their leading models, so independent verification is hard. But the claims are pointed.
Size is part of the story. Moonshot describes Kimi K3 as having 2.8 trillion parameters, that is, 2.8 trillion numbers the model learned during training that shape how it thinks and answers. Alibaba puts Qwen3.8 at 2.4 trillion. Parameters are a rough guide to scale, not a guaranteed measure of quality. A smaller, well-trained model can beat a larger one on many tasks.
What may matter more is what both companies plan to do with these models.
Why should ordinary people care about this?
If these models are released as promised, developers anywhere in the world can download them, run them on their own computers, and build products on top of them, for free. That changes who gets to build AI tools and at what cost.
Moonshot says it will publish the full model weights on July 27, 2025. Alibaba says Qwen3.8 is "going open-weight soon." Once those files are public, the models are out of any company's control.
This openness contrasts sharply with the leading US labs, which keep their most powerful models proprietary. The US government has also used export controls to restrict China's access to the advanced chips, called GPUs, that power AI training. Those restrictions have not stopped these releases.
First reported by The Verge AI, the announcements echo the shock caused by DeepSeek last year, when a Chinese lab released a capable model built at surprisingly low cost. That moment rattled the assumption that enormous US spending on chips and data centres guaranteed a lasting edge.
That question is very much open again.
