DT-Router

English | 简体中文

Quality-aware LLM routing, with cost in mind.

DTRouter is a proprietary LLM routing method developed by Ant Digital Technologies (蚂蚁数科).

Ant Digital offers two products for model services and task delivery: DTMaaS gives developers unified API access to multiple models for building AI applications; AiWork brings coding and office work into a workflow with shared context, autonomous planning, execution, and verification.

DTRouter explores the model-selection problem relevant to both model services and agent workflows: choosing a suitable model for each request's capability requirements and cost constraints, while balancing answer quality and inference spending. This repository presents that routing work and provides entry points to both products.

Explore MaaS · Try AiWork · MaaS documentation

1. Why model routing?

Different models offer different strengths, response times, and prices. Choosing a model for each request gives developers another way to manage the quality and cost of an AI application.

DTRouter explores this trade-off through request-level model selection. Its goal is to use model capabilities where they add value while keeping inference spending in view.

2. Build with Ant Digital MaaS

Interested in using multiple models in your applications? Explore DTMaaS, Ant Digital's enterprise Model-as-a-Service platform.

MaaS provides a unified API for accessing supported models, including LingDT, Qwen, and DeepSeek.

Capability	What it means for developers
Unified model access	Use OpenAI-compatible interfaces and switch supported models through request parameters.
Service scheduling and failover	Use the platform's scheduling and node failover mechanisms for model serving.
Inference cost optimization	Benefit from the platform's caching and routing optimizations.
Unified billing	Manage model consumption through a common service platform.

See the platform introduction and platform advantages for product details.

2.1 Try MaaS

1. Visit the MaaS platform and obtain an API key.
2. Follow the quick access guide and choose an available model from the supported models.
3. Send your first request through the OpenAI-compatible endpoint.

The example below shows how to call a specific model through MaaS.

curl https://maas-api.antdigital.com/v1/chat/completions \
  -H "Authorization: Bearer ${MAAS_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.6-plus",
    "messages": [
      {"role": "user", "content": "Explain the role of model routing in AI applications."}
    ]
  }'

For model access and product integration, start with the MaaS documentation.

3. AiWork: coding and office work together

Agent Team is Ready.

AiWork brings coding and office work into one workflow. Both modes share context, with autonomous planning, execution, and verification to move tasks toward deliverables you can inspect.

Mode	What AiWork helps you do
Coding	Develop maintainable code with attention to architecture and engineering quality. Check progress against the goal, diagnose issues, repair and revalidate, with reviews covering the plan, code, and overall result.
Office work	Read, convert, and summarize documents; research and analyze online data; operate your computer and web pages; connect tools through connectors and MCP; use expert skills; and schedule unattended tasks.

Stay involved throughout delivery. Users can approve, reject, or take over work while it runs. Outputs are traceable, and delivery is checked against verification evidence.

Use MaaS for model API access, and explore AiWork for coding and office tasks with shared context. See the AiWork homepage for the product overview.

Explore AiWork → · Download AiWork