Alibaba Targets 5–10 Trillion Parameters for Qwen's Next Generations

The roadmap from Apsara

At Alibaba Cloud’s annual Apsara Conference, CEO Eddie Wu put numbers on where the Qwen family goes next. The current flagship, Qwen 3.8 Max, runs at 2.4 trillion parameters. Qwen 4, still in training, will be larger. The Qwen 4.5 and Qwen 5 lines that follow are targeting 5 trillion to 10 trillion parameters.

Wu tied the work to tasks that are more complex and run over longer horizons, and described the goal as moving toward artificial superintelligence. He also said the team continues to research model architecture and data optimization.

The inference numbers matter just as much

Announcing parameters without the serving stack is an incomplete story, and Alibaba addressed both. Wu said the company’s in-house M890 AI supernode can already handle inference for models above 2 trillion parameters. Its successor, the Zhenwu V900, is claimed at three times the performance of M890, with mass production expected in the first quarter of 2027.

Put those two schedules side by side and the plan becomes legible: a 5–10 trillion parameter model and the V900 land in roughly the same window. Training at that scale is about cluster size and parallel efficiency. Serving it under real traffic is about something else — memory bandwidth, interconnect, and cost per token. Every step up in parameter count forces all of those numbers to be recalculated. That Alibaba is building the model and the inference silicon together is the most consequential part of the announcement, and the part least visible in a parameter headline.

Scale is not a capability claim

Parameter counts are the easiest thing to announce and among the easiest to challenge. As models grow, training cost, data quality, and benchmark saturation all get worse at the same time. Comparability is a separate problem: frontier systems are overwhelmingly sparse, so “total parameters” and “parameters active per token” have drifted apart. A number can be accurate and still not mean what readers assume.

The multiplier deserves to be stated plainly as well. Five to ten trillion is roughly two to four times the size of the current flagship, arriving only a few generations later. Unless architecture and efficiency work absorb part of the bill, that pace is not kind to a training budget.

So whether Qwen 5 at 5–10 trillion is a real jump or a figure that gets quoted for a year depends on architecture. The announcement did not include architecture details, a release date, training data scale, or whether 5–10 trillion refers to total or active parameters. There were no benchmark results either. Those gaps are exactly where to look when judging how much of this is a plan and how much is a direction of travel — the difference between a target and a shipped system.

What can be checked later

Three things are verifiable from here. First, whether Qwen 4 shows a measurable gain over Qwen 3.8 Max on public evaluations; a bigger model that does not move the benchmarks tells you something about where the returns are. Second, whether the Zhenwu V900 ships on the stated schedule. Alibaba already has a previous generation of inference hardware in service, so this is an execution question rather than a conceptual one, though 2027 is far enough out to leave room for slippage. Third, whether open-weight releases keep pace with the flagship line. Qwen’s developer base is built on downloadable checkpoints, and a roadmap slide does not fine-tune.

Alibaba has raised its own ceiling and put a date on the silicon. Everything else is a result that hasn’t arrived yet.