Alibaba plans a 10 trillion parameter Qwen and a chip to train it on
At Apsara in Hangzhou, Eddie Wu put a number on the next Qwen generation and unveiled an in-house accelerator that Alibaba says scales to 500,000 units.
Today's flagship Qwen runs about 2.4 trillion parameters. Alibaba says the next one is four times that.
At Alibaba's Apsara Conference in Hangzhou on September 22, 2026, chief executive Eddie Wu said the company plans to train an AI model of 5 trillion to 10 trillion parameters, and unveiled an in-house accelerator called the Zhenwu V900. Reuters reported both announcements from the stage, syndicated here via Investing.com.
The scale claim needs its baseline. Alibaba's current flagship Qwen runs at roughly 2.4 trillion parameters, so the stated target is two to four times larger. The Zhenwu V900 comes from T-Head, Alibaba's own semiconductor unit. Alibaba says it delivers three times the performance of its predecessor, the M890, that it scales to clusters of up to 500,000 units, and that mass production starts in the first quarter of 2027.
Behind both sits a power number. Alibaba is targeting 20 gigawatts of data center capacity by 2032, which is the figure that makes the other two plausible or not.
Announced, not benchmarked
Everything above is a company statement made at its own conference. No independent benchmark of the Zhenwu V900 exists. The 3x figure against the M890 is Alibaba's, measured by Alibaba, on workloads Alibaba chose. The model does not exist yet, so its parameter count is a plan, and parameter count has not predicted capability reliably for two years. The one checkable detail is the Q1 2027 production date, which either happens or does not.
This is the same pattern as Alibaba's 2.4 trillion parameter Qwen 3.8 preview in July, which arrived with no published benchmarks and then shipped in August with them. The preview turned out to be roughly honest. That is a reason to keep watching, not a reason to accept the next number in advance.
Why a build studio cares
Qwen is not an abstraction for anyone routing models. It is the open-weight line that keeps resetting the floor price of a capable model, and several of the releases we have covered this quarter were built on top of it rather than competing with it. What changes here is where the constraint sits. For two years the answer to "can we self-host this" was about weights and license. Alibaba is telling you the next answer is about silicon and gigawatts, neither of which ships to your cluster. A 10 trillion parameter model with permissive weights is still a model you rent, and the company that makes the chips and the power is the company that sets that rent.
Next step: read Reuters on the announcement, and put the Q1 2027 Zhenwu date in a calendar as the thing to check. If your product is priced on today's token costs, write to us at hello@gattyworks.com.