Model compression

​

Feature overview

Model compression uses quantization to convert a full-precision fine-tuned model into a low-precision version, reducing the MU specification required for deployment while preserving model capabilities, thereby lowering inference deployment costs. On the Model Studio platform, model compression specifically refers to quantization and does not involve structural pruning or knowledge distillation. Model compression is an optional step in the model production pipeline. The complete pipeline is: Fine-tuning → Model compression → Deployment. You must first complete fine-tuning training, then compress the resulting fine-tuned model, and finally deploy the compressed model as an inference service.

​

Supported models

The models currently supported for compression are listed in the table below. Refer to the console for the actual list.

​

Usage limits

• Available only in the North China 2 (Beijing) region.
• Only custom models produced through fine-tuning on the Model Studio platform are supported. Base models and third-party models are not supported.
• Compressed models cannot be further fine-tuned, nor can they be compressed a second time.

​

Create a compression task

​

Prerequisites

Before using the model compression feature, ensure that a custom model fine-tuned on the Model Studio platform is already available in the current workspace. If you have not yet completed model fine-tuning, first complete the training by referring to Fine-tuning.

​

Procedure

1. Log in to the Model Studio console. In the left navigation pane, select Model Compression.
2. On the model compression page, click Create Compression Task. The upper-right corner of the page also provides a Back to Old Version entry.
3. On the create compression task page, configure the following parameters:
   

   Parameter	Required	Description
   Task Name	Yes	Up to 50 characters. It is recommended to include the model abbreviation, quantization method, and version number for easier subsequent management.
   Task Description	No	Up to 200 characters.
   Source Model	Yes	Click to open the model selection dialog, which only displays fine-tuned models that can be compressed. If no available models appear in the dialog, complete model fine-tuning training first. Switching the source model automatically clears the selected quantization template, because the template lists supported differ across models.
   Output Model Name Suffix	Yes	Only lowercase letters and digits are supported, up to 8 characters in length. Invalid characters are automatically filtered out. This suffix is appended to the source model name as the identifier of the compressed model.
   Quantization Template	Yes	Available templates are displayed as a card grid. You must select a source model before selecting a template. A larger MU number in the template name indicates a smaller deployment specification and lower cost.
   Calibration Data	Conditionally required	This field appears only when the selected quantization template includes calibration input parameters. You can select up to 5 datasets; OSS-mounted datasets are not supported. Datasets must be created and published in Data Management.

4. After confirming the configuration is correct, click Start Compression. The button becomes clickable only after all required fields are filled in. Upon successful creation, you are automatically redirected to the task list page.

​

View compression task progress and logs

On the task list page, click a task name to enter the details page. The details page contains two tabs: Details and Log.

​

Details tab

The Details tab displays the following information:

• Basic information: Task ID (copyable), running status, creation time, and task duration (updated in real time while running). When a task fails, the error message is additionally displayed.
• Compression configuration: Source model, compressed model, quantization method, calibration data, and other configuration information set at creation time.

​

Log tab

Switch to the Log tab to view the running logs of the compression task.

• Lines displayed: You can choose to display the most recent 100, 500, 2000, or 5000 lines of logs.
• Auto refresh: When enabled, the log content refreshes automatically every 30 seconds.
• Log level coloring: INFO (green), DEBUG (gray), ERROR (red), WARN (yellow).
• Log operations: Supports downloading full logs (not limited by the displayed line count), collapsing or expanding sub-entries, switching between dark and light modes, and toggling word wrap.

​

Failure troubleshooting

If a compression task fails, troubleshoot by following these steps:

1. View the error message on the Details tab.
2. Switch to the Log tab and search for ERROR-level logs.
3. Download the full logs for detailed analysis.
4. If the issue remains unresolved, submit a ticket and attach the task ID and log file.

​

Manage compression tasks

In the console, select Model Compression in the left navigation pane to enter the task list page and manage the compression tasks that you have created.

​

Task list

The task list displays the following 7 columns of information:

Column name	Description
Task Name/ID	Task name and task ID. Click the task name to go to the details page.
Quantization Template	The name of the quantization template selected when the task was created.
Task Status	The current running status of the task.
Source Model	The name of the fine-tuned model before compression.
Compressed Model	The name of the model produced by compression. This is clickable only when the task succeeds, redirecting to the model center for viewing.
Creation Time	The creation time of the task.
Operation	The available operations, including stop and delete. Operation availability depends on the task status.

You can filter and search by Quantization Template, Task Status (5 types), time range, and task name or ID. The list supports sorting and displays 10 records per page.

​

Task status

Compression tasks have 7 statuses in total. The lifecycle is as follows:

• PENDING (not started): The task has been created and is waiting to be scheduled.
• QUEUING (in queue): The task has entered the execution queue.
• RUNNING (running): The task is executing compression.
• SUCCEEDED (compression succeeded): Compression is complete and the model has been produced.
• FAILED (compression failed): An error occurred during compression.
• CANCELING (stopping): The user initiated a stop and the task is being terminated.
• CANCELED (canceled): The task has been stopped by the user.

​

Stop and delete tasks

• Stop a task: Only tasks in PENDING and RUNNING statuses can be stopped. After clicking Stop, you must confirm again in the confirmation dialog. Once stopped, the task cannot be recovered. A stop button is also available on the details page.
• Delete a task: Only tasks in terminal states (SUCCEEDED, FAILED, CANCELED) can be deleted; tasks in the QUEUING status cannot be deleted. Deleting a task record does not affect the already-produced compressed model.

Empty state prompt: When there are no tasks, the page displays "You have not created any compression tasks yet" guidance text and provides a Create Compression Task button entry; when there are filter results, it prompts "No matches found", and suggests adjusting the filter conditions or clearing the search.

​

Billing

• The compression task itself is free for a limited time. The deadline is subject to console announcements.
• The compressed model is billed by MU specification during the deployment stage. For detailed deployment billing information, see Deployment.

Take the qwen3.5-flash-2026-02-23 fine-tuned model as an example. For the deployment specifications and prices before and after compression, see Supported models. Deployment costs are reduced by approximately 56%. We recommend comparing the compression effects of multiple quantization templates during the free period and selecting the optimal plan before formally deploying.

​

FAQ

What if no model is available when creating a compression task?

The source model selection dialog only displays fine-tuned models that can be compressed. If there are no selectable options in the dialog, it means that the current workspace does not have a qualified fine-tuned model. Please first refer to Fine-tuning to complete the fine-tuning training, and ensure the training status is successful before creating a compression task.

Can the compressed model be further fine-tuned or re-compressed?

No. Compression is a one-way operation. If you need to make adjustments, please re-compress from the upstream full-precision fine-tuned model. For details, see the usage restrictions in Feature overview.

What is the difference between quantization, pruning, and distillation?

These are three different model optimization techniques. Quantization reduces parameter precision, pruning removes redundant structures, and distillation uses a large model to train a small model. Model Studio currently only supports quantization.

How do I choose a quantization template?

The quantization template determines the deployment specification of the compressed model. The larger the MU number in the template name, the smaller the deployment specification and the lower the cost, but the greater the possible precision loss. We recommend choosing a template based on your trade-off between cost and precision for your business. If you are unsure, you can try multiple templates during the free period and verify the effects using a business test set after deployment.

How do I choose calibration data?

Calibration data is used to improve quantization precision. We recommend choosing a dataset that is semantically similar to the target inference scenario. For example, if the compressed model is used for a customer service Q&A scenario, we recommend choosing a dataset containing customer service dialogue samples as the calibration data.

What if the compressed model's precision drops?

Quantization compression may cause a certain degree of precision loss. We recommend using a business test set to perform inference verification after deployment. If the precision does not meet your needs, you can try changing the quantization template and re-compressing, or add calibration data to improve quantization quality.

How is billing handled after the free period ends?

Compression tasks are currently free for a limited time. The billing method after the free period ends is subject to console announcements. The deployment fee of the compressed model is not affected by the free period and is always billed by MU specification.

How do I use the free period to determine the best compression plan?

We recommend trying multiple quantization templates on the same fine-tuned model during the free period, deploying them separately, and verifying the inference effects using a business test set. Compare the performance of different templates between precision and cost, and select the plan that best fits your business needs before going live formally.

If you need to manage compression tasks via API, see Model compression API reference.