1. Home
2. News
3. Open weights

Alibaba says Qwen ran a month of self-improvement unattended, and plans a 10-trillion-parameter model

Eddie Wu's keynote says the Qwen team has made meaningful progress on AI that improves itself, the trend Dario Amodei wants slowed.

By Think Facility · September 23, 2026 · Alibaba AI models

On September 22, 2026, Alibaba's CEO Eddie Wu opened the Apsara Conference in Hangzhou by saying the Qwen team plans to train a new model at the scale of 5 to 10 trillion parameters. Alibaba's release puts that size in the coming Qwen 4.5 and Qwen 5 series. Qwen 4 is in training now.

The same keynote says the Qwen team is exploring recursive self-improvement and has made meaningful progress. I didn't expect a lab to put that on stage ten days after the pacing debate began, since it's the trend Dario Amodei gave as his first reason for asking the labs to slow down.

Who

Eddie Wu, CEO of Alibaba Group

Where

2026 Apsara Conference, Hangzhou

Next models

Qwen 4.5 and Qwen 5, 5 to 10 trillion parameters

Self-improvement run

33 iterative cycles over a month

New chip

Zhenwu V900, mass production in Q1 2027

What Alibaba says Qwen did on its own

The press release puts numbers on it. Over a month of fully automated runs, Qwen3.8-Max went through 33 cycles of pipeline design, data validation, experiments and error diagnosis, and its Artificial Analysis score went from 40 to 45.

A second run pointed the model at chip design. It spent more than 60 hours improving itself across the design lifecycle and made more than 10,000 EDA tool calls. Alibaba says the chip bus modules it produced cut chip area by 42% with no loss of performance.

The release doesn't say who checked the 40-to-45 score, or whether the updated Qwen3.8-Max is the one customers use. In the keynote, Wu says machine thinking will ultimately do 99.9% of all thinking.

Next to the pacing essay

Amodei's essay, We Must Pace the Frontier, says recursive self-improvement is starting to happen across the industry. One of the steps he proposes is a speed limit on it.

Wu's keynote doesn't mention slowing down. He calls RSI a concrete path forward that AI researchers have mapped out, and says the bigger model is meant to handle longer tasks and advance toward ASI.

The chip under it

The hardware is the Zhenwu V900 from T-Head, Alibaba's chip design subsidiary. Wu called it the most powerful AI chip in China today. It has three times the performance of the Zhenwu M890 that came out in May, and a single cluster can take up to 500,000 cards.

The V900 is scheduled for mass production and commercial release in Q1 2027. Wu's target for Alibaba Cloud's data center capacity is more than 20GW by 2032.