Alibaba wants to supply the chip, model and cloud. What choices remain?

A company chooses a Qwen model to answer employees’ questions. Months later, the application also depends on a fine-tuned model, document retrieval and access rules. Changing providers now means rebuilding that behavior, even if the original model remains available to download.

This is a useful way to examine Alibaba’s Apsara strategy. The group wants to bring chips, models and agent services closer together. A customer can delegate more operations to one provider. Its ability to leave depends on what it can recover at each layer.

Announcements and available products have different timelines

Alibaba’s September 25 conference recap introduces T-Head’s Zhenwu V900 with mass production and commercial release planned for the first quarter of 2027. Customer references concern the Zhenwu chip family; they do not establish that those customers already use V900. Official Apsara recap.

The ambition also covers training, agent deployment and context management, the information available to agents when they act. It does not establish that every Qwen API call runs on T-Head hardware or that every service is available in every region.

A downloadable model, a documented feature and an announced chip do not necessarily form an offering that the same customer can purchase immediately.

Choosing Qwen leaves several deployment paths

Qwen3-14B illustrates why choosing an Alibaba model does not require choosing its entire cloud. The Qwen3 project publishes open-weight models and documents deployment with runtimes including vLLM and SGLang. Weights are the parameter files needed to execute the model. Official Qwen3 project.

A company can assemble its service on infrastructure it administers. In that scenario, it retains model files, the code coordinating tools, documents and evaluation tests. It also takes responsibility for upgrades, monitoring and incidents.

That control requires ongoing work. Possessing weights does not reproduce an application if runtime versions, settings or document processing have been lost. Changing numerical precision or retrieval behavior can change responses even when the model name stays the same.

Managed services offer another route. They can reduce the infrastructure a team operates itself, while adding resources whose recovery depends on the provider’s interfaces.

In Model Studio export does not mean download

Model Studio’s fine-tuning guide provides a concrete example. Fine-tuning adapts a model using customer examples. A checkpoint records a state of its parameters. In the documented public workflow, the operation called export prepares that checkpoint for deployment within the service. The guide states that exported checkpoints remain in cloud storage and cannot be downloaded. Model Studio guide, model publishing.

This restriction applies to that workflow. It does not establish that every Alibaba service prevents model recovery, or rule out different terms under a particular agreement.

For the document assistant, the consequence is tangible. If its behavior depends on that adaptation, downloading the original public model does not recover the subsequent work. Retaining the training examples could make a new adaptation elsewhere possible, without guaranteeing identical results.

API compatibility solves a different problem. It can make sending requests from an existing client easier. It does not supply customized parameters or every state accumulated by the service.

Documents are only part of retrieval

A document assistant retrieves relevant passages before sending them to the model. This approach, retrieval-augmented generation or RAG, depends on document splitting, indexing and the rules selecting results.

Model Studio’s knowledge-base documentation says one-click export is not supported. It proposes retrieving documents and chunks through a script using ListChunks. That recovery path exists, but it does not describe a package restoring the whole retrieval service identically. Migration FAQ.

If a new engine splits documents or ranks passages differently, the model receives different context and may answer differently. Access rules also need to be restored. A recovered document must not become visible to an employee who was previously unable to read it.

| Application element | Documented fact or scenario condition | Consequence when leaving | |---|---|---| | Public model retained by the company | Files exist independently of its application | Redeploy with controlled versions and environment | | Model Studio adaptation in the examined workflow | Guide excludes checkpoint download | Obtain another recovery path or consider training again | | Managed knowledge-base documents and chunks | Scripted retrieval is documented | Rebuild and test retrieval behavior | | Company-owned code, rules and tests | An assumption of this scenario | Replace integrations without losing acceptance criteria |

The consequences in this table are architectural analysis, not results from an observed customer migration.

Other Alibaba services let customers retain files

The chosen service changes the assessment. PAI documents training and development environments that read and write OSS, Alibaba’s object storage. Files the company retains there follow a different path from Model Studio’s internal checkpoints. PAI and OSS documentation.

A single portability score for Alibaba would conceal this distinction. A company can rent compute while retaining its artifacts, or delegate more of the application. Its exit cost depends on that division of responsibilities.

Operational benefits belong in the comparison too. A managed service can spare a small team from maintaining several components. Replacing those components during migration takes work. Without measuring both efforts for the same need, neither approach can be declared universally cheaper.

Choice depends on what the application accumulates

Alibaba’s planned integration could simplify buying and operating a complex technical stack. An open model establishes only part of customer freedom. Adaptations, indexes, rules and accumulated state determine what must be reconstructed elsewhere.

For a company considering these services, an exit test on a small application would reveal more than a compatibility label. Recover the available artifacts, rebuild the workflow and rerun the same tests. That would distinguish what transfers, what needs conversion and what must be recreated.

Alibaba wants to provide more layers of AI infrastructure. The customer’s remaining choice will depend on whether it can move the service it built, including its behavior and rules, beyond the model file alone.

Sources and method

Documentary analysis prepared on October 6, 2026, using the October 2 dossier and targeted checks of primary sources. The document assistant is an explanatory scenario. No paid account, customer deployment or export was tested. Migration consequences are technical inferences, not allegations of abusive conduct.

Primary sources are the September 25 Apsara recap, Qwen3 project, Model Studio fine-tuning guide, knowledge-base FAQ and PAI/OSS documentation. Recovery restrictions remain attached to the cited workflows. No quantified cost advantage is asserted without a complete comparison at equivalent quality and service requirements.