Large model service overview
DataWorks large model services provide one-stop capabilities for efficient deployment, secure communication, and convenient invocation. You can easily deploy models on DataWorks serverless resource groups and call them directly in data development tasks. All traffic is transmitted through a private network channel established by PrivateLink, ensuring that data never leaves the private network and effectively safeguarding data security.
Supported models
Category | Model | Description |
Large language model | Qwen3-32B | Qwen3 is the latest generation of large language models in the Qwen series, offering a full set of dense and Mixture-of-Experts (MoE) models. Based on extensive training, Qwen3 achieves breakthrough capabilities in reasoning, instruction following, agent functions, and multilingual support. For more information about the model, see Recommended models. Important
|
Qwen3-14B | ||
Qwen3-8B | ||
Qwen3-4B | ||
Qwen3-1.7B | ||
Qwen3-0.6B | ||
Qwen3-Embedding-8B | ||
Qwen3-Embedding-4B | ||
Qwen3-Embedding-0.6B | ||
DeepSeek-R1-0528-Qwen3-8B | DeepSeek extensively uses reinforcement learning during training, significantly improving the model's reasoning capabilities with minimal labeled data. For more information about the model, see Recommended models. | |
DeepSeek-R1-Distill-Qwen-1.5B | ||
DeepSeek-R1-Distill-Qwen-7B | ||
DeepSeek-R1-Distill-Qwen-14B | ||
Embedding model | BGE-M3 | BGE-M3 is a general-purpose embedding model that supports dense retrieval, multi-vector retrieval, and sparse retrieval. It can process inputs ranging from short texts to a maximum of 8,192 tokens and supports more than 100 natural languages. |
BGE-Large-zh-v1.5 | BGE Embedding is a general-purpose embedding model. It uses RetroMAE for model pre-training, followed by contrastive learning on large-scale paired data. |
Large model service features
Quick deployment: Simplified launch process
With a visual interface guide, you can deploy pre-trained models with low-code operations. You do not need to write infrastructure code. With just a few simple configuration steps, you can create and publish model services, lowering the barrier to model deployment.
Secure private network communication: Ensuring service isolation
DataWorks large model services are deployed in fully managed mode on DataWorks serverless resource groups. Cross-VPC private network communication is established based on PrivateLink and Private Hosted Zone to achieve network isolation. You can access the fully managed model through your VPC private network to ensure the security of model services.
Network-isolated communication: PrivateLink establishes a dedicated connection channel between your VPC and the DataWorks resource group VPC, ensuring that traffic is not exposed to the public network and achieving secure isolation.
Private domain name resolution: Private Hosted Zone deploys a private DNS resolution service that automatically configures private domain name forwarding rules, allowing you to access model services directly through domain names within your VPC.
Automated connection: Seamless invocation
When you deploy a model service on a DataWorks serverless resource group or configure a VPC for it, the system automatically performs the following operations:
Establishes cross-VPC connections. In your VPC (the VPC under your account that is connected to the DataWorks resource group), the system automatically creates a PrivateLink endpoint and establishes an encrypted communication channel with the PrivateLink service in the DataWorks resource group VPC.
Configures domain name resolution. Domain name resolution rules are automatically configured in the VPC specified in the model service configuration, so that domain name request traffic within the VPC is automatically forwarded to the DataWorks model service.
Model invocation process
When you initiate a model service call through a domain name:
Request resolution: DNS requests within your VPC are resolved by Private Hosted Zone to locate the private IP address of the DataWorks resource group VPC.
Traffic forwarding: Requests are securely transmitted from your VPC to the DataWorks resource group VPC through the dedicated channel established by PrivateLink.
Service response: Requests reach the model instance in the DataWorks resource group VPC, where inference computation is completed and results are returned.
The entire invocation process runs within a private network environment without exposing public IP addresses, thereby ensuring the security and controllability of model services.
Model deployment and usage
In DataWorks, you can select and deploy models for task development.
Third-party models
In addition to deploying pre-trained models, DataWorks large model services also support registering third-party model services. You can connect APIs from external model providers (such as Bailian) to DataWorks. After registration and activation, you can call third-party models in a unified manner within DataWorks tasks.
On the Model Service page, switch to the Third-party models tab. You can:
Register a model service: Provide information such as the model provider, model name, service name, and API key to complete registration.
Manage model services: Enable, disable, modify, or delete registered third-party model services.
Third-party model services do not consume resources from DataWorks serverless resource groups. Actual costs are charged by the corresponding model provider.