At its annual Yunqi Conference in Hangzhou on September 22, Alibaba dropped one of the biggest AI announcements of the month: its next flagship model, Qwen4, has entered training — and it's "coming very soon."
The headline: 10 trillion parameters
Alibaba's roadmap is staggering. After Qwen4, the company says Qwen4.5 and Qwen5 will scale to somewhere between 5 trillion and 10 trillion parameters. For context, today's biggest models are measured in the low trillions at most. The scale race just entered a whole new bracket.
We already have a preview
Alibaba gave developers an early look at Qwen4's architecture on August 28 with Qwen 3.8-Flash-Next, described as "an experimental preview of the architecture that will underpin Qwen4." The design direction: a sparse mixture-of-experts model that activates only a fraction of its parameters per token, a large memory layer that trades computation for memorized lookup, native image and video input, and context windows up to 1 million tokens. In plain terms: leaner, cheaper to run, and multimodal by default.
The wild detail: an AI that did its own homework
One report from the conference floor is hard to ignore. Alibaba described a test where Qwen3.8-Max practiced recursive self-improvement — for more than a month, the model designed its own training process and generated its own data without human intervention, completing 33 training iterations. In one chip-design experiment, it invoked chip-design tools over 10,000 times and shrank chip area by 42%. Take the numbers as reported, but the direction is clear: models are starting to do some of their own homework.
Not just models — the whole stack
Alibaba also introduced Qwen Intelligence, a smartphone agent platform that handles task planning, cross-app execution, and video creation for handset makers. The Honor Magic9 will be the first phone to ship with it. Meanwhile the Qwen open-source ecosystem keeps growing: 3 billion cumulative downloads, 300,000 derivative models, and 460 open-source releases. On the same day, Xiaomi released its MiMo-V2.6 Pro and Flash models under the MIT license — China's open-source AI push is in full gear.
Why it matters
Chips, cloud, models, and now phones — one company controlling the whole stack is China's playbook for AI, and it's a different game than the US labs are playing. If Qwen4 lands with strong open weights like its predecessors, it'll put serious pressure on Western models in developer adoption worldwide.
My take
Don't sleep on Alibaba. While everyone's watching OpenAI and Anthropic trade price cuts, Qwen has quietly become the most downloaded open model family on earth — and a 10-trillion-parameter open model could change the conversation completely.
Comments
Post a Comment