Deploy and train models quickly-Platform For AI(PAI)-阿里云帮助中心

Model Gallery offers a variety of pre-trained models to help you get started quickly with model deployment and training.

Choose a model that fits your business needs

Model Gallery provides a wide range of models to help you solve real-world business problems. Use the following guidance to find the model best suited for your needs:

Search by domain and task: Find models based on your application domain and the specific task you want to perform.
Review pre-training datasets: Most models list the datasets used for pre-training. The closer the pre-training dataset is to your actual use case, the better the performance when deploying or fine-tuning the model directly. You can find more information about pre-training datasets on the model details page.
Consider model size: Generally, models with more parameters deliver better performance. However, they also incur higher costs during model serving and require more data for fine-tuning.

Follow these steps to find a suitable model:

Go to the Model Gallery page.
1. Log on to the PAI console.
2. In the navigation pane on the left, click Workspaces, then select and enter your target workspace.
3. In the navigation pane on the left, click QuickStart > Model Gallery to open the Model Gallery page.
Find a model that fits your business needs.

Under the Getting Started section in the PAI console’s left-side navigation pane, click Model Gallery to go to the Model Gallery page. Use the search box, filter by model source (ModelScope, PAI, NIM), or browse available model cards by scenario tags.

After selecting a model, you can deploy it directly and perform online debugging to validate its inference performance. For details, see Deploy models and Fine-tune models.

Deploy models

For detailed instructions using the Qwen3-0.6B model as an example, see Model Gallery Quick Start – Model deployment.

Fine-tune models

For detailed instructions using the Qwen3-0.6B model as an example, see Model Gallery Quick Start – Model fine-tuning.

On the fine-tuning job details page, configure the following parameters.

Fine-tuning configurable parameters

Parameter type	Parameter	Description
Training Mode	SFT supervised fine-tuning	Supported training methods include the following: Supervised fine-tuning: Fine-tune a Large Language Model (LLM) using input-output pairs. Direct preference optimization: Optimize a language model directly to align with human preferences, implicitly sharing the same objective as RLHF algorithms. Both methods support full-parameter fine-tuning, LoRA, and QLoRA.
Training Mode	DPO direct preference optimization
Job Configuration	Task name	A default job name is provided. You can change it based on the interface prompt.
Job Configuration	Maximum running time	Set the maximum allowed runtime for the job. If the job exceeds this duration, it stops automatically. If you keep the default setting, the job runtime is not limited by this parameter.
Dataset Configuration	Training dataset	Model Gallery provides default training data. If you do not use the default dataset, prepare your training data in the format specified in the model documentation, then upload it using one of the following methods. OSS file or directory Click , then select the OSS path where your dataset is stored. In the Select OSS folder or file dialog box, choose an existing data file or Upload file. Custom Dataset You can use datasets stored in cloud storage services such as OSS. Click to select an existing dataset. If you do not have a dataset, create one by following the instructions in Create and manage datasets.
Dataset Configuration	Validate dataset	Click Add validation dataset to add one. Configure it the same way as the Training dataset.
Output Configuration		Select a cloud storage path to save the trained model and TensorBoard log files. Note If you configured a default OSS storage path for your workspace on the workspace details page, this field is auto-filled and requires no manual setup. For instructions on configuring workspace storage paths, see Manage workspaces.
Computing Resources	Resource Type	Supports general computing and Lingjun Intelligent Computing.
Computing Resources	Source	Public Resources: Billing method: pay-as-you-go. Scenarios: Public resources may experience queuing delays. Use them when you have relatively few jobs and low time sensitivity. Resource Quota: Includes general computing or Lingjun Intelligent Computing resources Billing method: subscription. Scenarios: Ideal for high-volume workloads requiring guaranteed execution. Preemptible Resources: Billing method: pay-as-you-go. Scenarios: Use preemptible resources to reduce costs—they typically offer significant discounts. Limits: Availability is not guaranteed. Resources might not be immediately available or could be revoked. For more information, see Use preemptible jobs.
Hyperparameters		Different models support different hyperparameter configurations. Use the default values or modify them as needed.

Note

Available parameters vary by model. Configure them based on your specific model.

Billing

Model Gallery is free, but model deployment and training incur charges from Elastic Algorithm Service (EAS) and Deep Learning Containers (DLC). For details, see Elastic Algorithm Service (EAS) billing and Deep Learning Containers (DLC) billing.

References

Model Gallery FAQ