Alibaba's Apsara roadmap: 10 trillion parameters is a projected range, not an imminent release

Why it matters

【Key points】

At its Sept 22 Apsara Conference, Alibaba put models, chips and data centers on one roadmap. Qwen 4 "is currently in training"; the Qwen 4.5 and Qwen 5 series are "projected to scale up to 5 to 10 trillion parameters." None has a release date. T-Head's Zhenwu V900 accelerator is "scheduled for mass production and commercial release in Q1 2027," and CEO Eddie Wu targets over 20GW of Alibaba Cloud data center capacity by 2032. Editorial synthesis from Alibaba's releases.

【Verifiable】

Press release (Sept 22): https://www.alibabacloud.com/en/press-room/alibaba-unveils-roadmap-on-full-stack-ai-strategy ; recap (Sept 25): https://www.alibabacloud.com/blog/603598 . V900: 3x the Zhenwu M890 (released in May), 216GB GPU memory, 1,200GB/s inter-chip bandwidth, FP8/FP4. A V900 supernode server "can support a supernode cluster comprising up to 500,000 cards" (no date). Self-improvement: over a month of fully automated runs, Qwen3.8-Max "completed 33 iterative cycles" and its "Artificial Analysis score" rose "from 40 to 45." Alibaba doesn't say which metric or version. In August Artificial Analysis put Qwen3.8 Max at 56 on its Intelligence Index (https://x.com/ArtificialAnlys/status/2085270415614828675 ); we found no evaluation of the self-improved version.

【Our read】

1. The story isn't "10 trillion" but the coupling. Alibaba names no training chip for Qwen 4.5/5; if they rely on V900 clusters, its own timetable means no large-scale training before V900 ships in 2027. That is our inference, not Alibaba's claim.

2. "A 10-trillion model is coming" misreads an undated projected range for two future series. The near-term test is Qwen 4: when it ships with a model card, pricing and an API.

3. "33 cycles, 40 to 45" is self-reported, with no third-party check.

4. Chinese X complaints that domestic labs have gone quiet don't contradict Alibaba's schedule, which only says Qwen 4 is training. We expect no 5T+ Qwen before V900 volume production.