Alibaba Announces the Zhenwu V900 Chip and Expands Qwen’s Roadmap: What’s Available and What’s Still a Plan

What Alibaba announced at Apsara

On September 22, 2026, Alibaba announced an update to its artificial intelligence strategy spanning chips, cloud infrastructure, and models. The main announcements included the Zhenwu V900 processor, the development status of Qwen 4, and projected scale for the future Qwen 4.5 and Qwen 5 series. These are different kinds of developments: a chip announced with a target commercialization date, a model the company says it is training, and development goals for later generations.

That distinction matters to anyone assessing technology for immediate use. A product announcement does not mean general availability; a roadmap projection is not a specification for a model that has already been released; and a manufacturer’s performance claim is not a substitute for a comparable test. Alibaba’s announcement and Reuters coverage republished by MarketScreener provide information about the announcements, but the material available for this article does not include independent tests validating the chip’s performance figures.

Official documentation also provides context for Qwen3.8-Max as an existing model, distinct from the announcements about Qwen 4 and its successors. The model information page identifies it as a system with 2.4 trillion parameters. That figure offers context, but is not enough on its own to establish the model’s quality, cost, or performance on a particular workload.

Zhenwu V900: announcement, performance, and target date

Alibaba introduced the Zhenwu V900 and said it delivers three times the performance of the Zhenwu M890. This comparison should be attributed to the company. The material available for this article does not specify the metric used, the chip configurations, the workload, or the testing conditions. Without those details, it is not possible to fully interpret what “three times” means or to compare the result with products from other manufacturers.

The company set mass production and commercial launch of the V900 for the first quarter of 2027. This is a target date, not confirmation that the product can already be purchased or installed. Nor does the available information independently verify whether the schedule will be met or what access, pricing, and regional availability terms will apply.

For developers and infrastructure decision-makers, physical availability is only one part of the evaluation. Memory, bandwidth, supported precisions, drivers, development tools, and compatibility with existing software also matter. The sources provided do not confirm these technical details for the V900. It is therefore best not to infer them from the performance announcement or target date.

Qwen 4 is in training; the 5-to-10-trillion figure is a projection

Alibaba said Qwen 4 was in training. That description refers to the development status the company reported for the model; it does not establish that the model is available or confirm a release date in the sources provided. The announcement also assigns a projection of 5 to 10 trillion parameters to the future Qwen 4.5 and Qwen 5 series. It does not present that figure as the size of Qwen 4.

In English, a trillion is 10¹². Even with the scale clarified, the number of parameters does not automatically reveal a system’s practical capabilities. Evaluating a model also requires, among other things, results on relevant tasks, an evaluation methodology, inference conditions, cost, and availability. The sources available here do not support turning the projected figure into a conclusion about quality or usefulness.

Keeping the generations separate prevents a misreading of the announcement: Qwen 4 is the model the company says is in training; Qwen 4.5 and Qwen 5 are the future series associated with the projected scale. The sources provided do not support claims that those versions have already been released, that their final size has been settled, or that the targets have been achieved.

Qwen3.8-Max as an available reference, not proof of future announcements

Alibaba had already introduced Qwen3.8-Max, and its Model Studio documentation identifies it as a model with 2.4 trillion parameters. It is useful as a reference point for a product documented by the company, but should not be confused with Qwen 4 or the projections for Qwen 4.5 and Qwen 5. Documentation for a model is also different from an external evaluation of its results.

The information verified for this article does not provide an independent measurement that would confirm Qwen3.8-Max’s size or results beyond the figures attributed to Alibaba. It also does not include reproducible comparisons between this model and systems from other providers. Reuters coverage republished by MarketScreener is a journalistic source on the Apsara announcements, but does not, by itself, constitute a technical test of the chip or the models’ capabilities.

For that reason, this article distinguishes between information documented by Alibaba and independent validation. Official documentation is a relevant source for describing product availability or specifications; assessing comparative capabilities would require tests with transparent tasks, configurations, and conditions. The available sources do not establish that such an independent evaluation has been published.

What developers and buyers should check

When deciding whether to adopt a platform, it is most useful to turn broad announcements into verifiable questions. For the V900, teams will need to consult public technical documentation when it becomes available, including details on memory configuration, bandwidth, supported precisions, tool support, and compatibility with the required frameworks. They should also check which regions and channels will offer the chip and whether the announced timeline remains on track.

The practical questions for models are different: when they will be released, what access options will be available, whether weights or endpoints will be offered, what the prices and requirements are, and how the models perform on representative tasks. The sources provided do not confirm the publication of weights, endpoints, prices, or access requirements for the models announced at Apsara. It is not prudent to assume that they will be available simply because the company has announced a roadmap.

For infrastructure procurement, a useful comparison should reflect an organization’s actual workloads, not just a summarized performance figure. It is worth measuring latency, throughput, total cost, and stability under equivalent conditions, as well as reviewing operational support. Until documentation and reproducible tests are available, the V900 announcement helps explain Alibaba’s strategic direction, but is not enough on its own to justify a migration.

What can—and cannot—be concluded yet

The Apsara announcements set out a roadmap: Alibaba introduced the Zhenwu V900, placed its mass production and commercialization in the first quarter of 2027, said Qwen 4 was in training, and linked a future scale of 5 to 10 trillion parameters to Qwen 4.5 and Qwen 5. The V900’s comparison with the M890 is a company claim; the sources examined here do not provide enough methodology to reproduce it.

Based on the available information, it cannot be concluded that the V900 is already commercially available, that it will meet the target date, or that its claimed advantage will hold for workloads other than those Alibaba may have used. Nor should the parameter projection be treated as a final specification for Qwen 4 or as direct evidence of capability.

The most accurate way to follow these announcements is to keep documented products, manufacturer statements, and future goals separate. As technical manuals, access terms, and reproducible evaluations become available, it will be possible to assess more confidently which options are useful to particular teams. Until then, Alibaba’s roadmap describes its plans, but leaves important questions about performance, access, and real-world use open.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.