Alibaba plans 10-trillion-parameter Qwen models as part of a broader push spanning software, custom chips and data centers. Chief executive Eddie Wu presented the roadmap at the Apsara Conference in Hangzhou on September 22, pairing the model target with new details about the company’s Zhenwu V900 AI accelerator.

 

The plan would take Alibaba well beyond its current 2.4-trillion-parameter flagship, but the company has not released benchmarks, training costs or an exact launch date for the largest systems. The announcement therefore establishes Alibaba’s intended scale, not the eventual quality or commercial impact of the models.

 

Alibaba disclosed three connected parts of the expansion:

  • Qwen 4 is already in training.
  • Qwen 4.5 and Qwen 5 are projected to reach 5–10 trillion parameters.
  • Cloud capacity is targeted to exceed 20 gigawatts by 2032.

 

More on This Story

 

Alibaba Plans 10-Trillion-Parameter Qwen Models

Wu said Alibaba intends to train future Qwen systems with between 5 trillion and 10 trillion parameters. Qwen 4 is the immediate model under development, while Qwen 4.5 and Qwen 5 are the generations expected to enter that larger range.

 

The comparison point is Qwen 3.8-Max, released in August with 2.4 trillion total parameters and 95 billion active parameters. Alibaba describes it as its most capable Qwen model to date. The proposed upper end for Qwen 5 would be more than four times larger by total parameter count.

 

That increase is designed for complex, long-horizon work rather than shorter question-and-answer exchanges. Alibaba has been emphasizing agents that can plan, use tools and revise their work over extended sessions, and Wu said the company has made progress on systems that improve through repeated feedback.

 

The company’s public roadmap remains incomplete. Alibaba did not give the architecture, active-parameter count, training-compute budget, dataset composition or deployment price for Qwen 4.5 and Qwen 5. Those omissions make it impossible to compare efficiency with competing frontier systems before independent evaluations are available.

 

Zhenwu V900 Expands Alibaba’s Chip Roadmap

Alibaba also detailed the Zhenwu V900, a next-generation AI accelerator designed by its T-Head semiconductor unit. The company says the chip delivers three times the performance of its M890 predecessor and is scheduled for mass production and commercial release in the first quarter of 2027.

 

The performance figure is an Alibaba claim, and the company did not publish the benchmark methodology, power draw or memory configuration needed for an independent comparison. Those specifications will matter because frontier model training depends on memory bandwidth, networking and software support as much as headline compute throughput.

 

Alibaba said one cluster based on its hardware can scale to as many as 500,000 cards. Its existing M890 supernode is already used for inference on models exceeding 2 trillion parameters, creating a bridge between the present Qwen generation and the much larger systems on the roadmap.

 

Custom silicon gives Alibaba another path around constrained access to leading imported accelerators. U.S. export controls have limited shipments of advanced AI chips to China, encouraging Chinese cloud providers and chip designers to assemble domestic alternatives for training and inference.

 

Alibaba Targets 20 GW of Cloud Capacity

The model and chip plans sit inside a substantial infrastructure expansion. Alibaba Cloud is targeting more than 20 gigawatts of global data-center capacity by 2032, a measure of the electrical scale available to power servers, cooling systems and supporting equipment.

 

Wu said AI-compute demand continues to exceed supply, while Alibaba expects its supernodes to begin commercial-scale operation during the current quarter. The combination of tight near-term capacity and an aggressive long-term target shows that the company expects model deployment to become an infrastructure business, not merely a software release cycle.

 

Reaching 20 GW would require sustained investment in power procurement, networking, cooling and regional data-center construction. It would also expose Alibaba to the same questions facing other hyperscalers: whether grid connections can arrive on schedule, whether customers will use the installed capacity and whether efficiency improvements can offset rising electricity demand.

 

The vertically integrated strategy links each layer. Larger Qwen models create demand for training and inference; V900 accelerators provide a proprietary compute platform; and expanded cloud sites give Alibaba a route to sell that capacity to external developers and enterprises.

 

Parameter Scale Is Not a Performance Guarantee

Parameter count is a rough measure of model size, not a direct score of intelligence or usefulness. Architecture, training data, active parameters, post-training methods and inference-time computation can allow a smaller system to outperform a larger one on specific tasks.

 

Alibaba’s current Qwen 3.8-Max illustrates that distinction: its mixture-of-experts design activates only a fraction of its 2.4 trillion parameters for each token. Future models could use a similar approach, but Alibaba has not disclosed enough technical detail to determine their likely operating cost.

 

The next verifiable milestones will be model evaluations, detailed V900 specifications and evidence that the promised supernode deployments are operating at scale. Until then, the Apsara announcement is most significant as a declaration that Alibaba intends to compete across the entire frontier-AI stack rather than rely on one breakthrough product.