Alibaba Cloud Bailian: One Platform for Model Serving and Agent Building

Building a large-model application usually means solving four separate problems: calling a model, wiring an agent, managing a knowledge base, and staying stable under load. Alibaba Cloud Bailian puts them in one place, with a model marketplace for selection, templates and visual workflows for assembly, and reserved throughput plus batch jobs for cost and stability in production.

Bailian is Alibaba Cloud’s platform for model services and application development. It combines inference, agent development, RAG knowledge bases and workflow orchestration, aimed at enterprises and developers working on customer service, e-commerce marketing, short drama production and office automation.

The model layer covers more than a hundred multimodal models: the Qwen3.8 family, Wan3.0 for video, DeepSeek, plus third-party options such as Kimi-K3 and GLM-5.2. A single API key reaches all of them, which saves the work of integrating each vendor separately when a team wants to compare price and quality per task.

The engineering layer is about getting to production. Fine-tuning adjusts a general model to business wording, dedicated and reserved throughput handle high concurrency and low latency, and caching with batch inference lowers the cost of work that does not need an immediate answer.

Two layers sit around the core: a security programme covering infrastructure, models and applications, documented in a published whitepaper, and an application marketplace collecting models, tools and partner solutions.

Main capabilities

Model serving

Over a hundred multimodal models behind one interface, with caching and batch inference for lower cost. Caching suits repeated questions, batch suits offline processing.

Training and tuning

Fine-tuning aligns output with business requirements, while dedicated and reserved throughput protect response times at peak load.

Agent development

Templates, RAG knowledge bases, visual workflows and hosted MCP servers let a team chain retrieval, reasoning and actions into a working agent.

CLI tooling

One-command API key setup lets an agent call more than ten model capabilities from a local development environment.

Scenario templates

Prebuilt solutions for voice of customer, office work, e-commerce marketing and short drama let a team ship something running before customising.

Security and marketplace

A full-chain MaaS security approach with a public whitepaper, plus a marketplace of models, tools and partner offerings.

How to get started

1. Register an Alibaba Cloud account and complete identity verification, which is required before any service can be enabled.
2. Open the Bailian console and activate the model service.
3. Create an API key in the console, or review the Token Plan first if subscription quota fits better.
4. Pick a model by task in the marketplace, for example a reasoning model, Wan3.0 for video, or a vision-language model for image input.
5. Test prompts in the playground before writing code, then call the SDK or HTTP endpoint with the key, model name and prompt.
6. Assemble the first agent from a template, a knowledge base or a visual workflow, then tune with fine-tuning or a different model size.
7. At launch, reserve throughput for peaks and keep costs down with usage monitoring plus batch and caching.

Plans and quota

• New users receive a free quota; the exact amount and validity are shown in the console.
• Pay-per-call billing and a subscription Token Plan coexist, with unit prices varying by model size and by whether caching or batch is used.
• Reserved and dedicated throughput are resource commitments billed separately from per-call usage and should be sized to peak concurrency.
• No complete price list is published, so budgets should follow the console bill and a commercial quote.

Who it suits

• Engineering teams connecting models to existing business systems.
• Developers who want to compare several vendors without separate integrations.
• Product teams building customer service, marketing, video or office scenarios.
• Teams with explicit concurrency and latency targets that need throughput guarantees.
• Organisations that need a documented security posture before rollout.

Limits to know first

• The source does not publish a full price list, so model rates, cache and batch discounts and throughput reservation costs must be checked in the console.
• Free quota is bounded in size and time and converts to paid usage afterwards, so budget alerts should be set before trials.
• Model versions move quickly; pinning an explicit version is safer than assuming the latest one.
• Fine-tuning and reserved throughput are advanced features that consume quota and engineering effort rather than working on activation.
• Throughput figures quoted in comparisons reflect specific plan tiers; real limits depend on region and account.
• The platform does not govern your business data by default, so knowledge base permissions, masking and log retention remain your design work.

FAQ

How is this different from calling a model API directly?

A direct API solves inference only. Bailian adds model selection, agent assembly, knowledge bases, workflows and production operations, removing the need to build an orchestration layer.

Do I need to write code?

Not necessarily. Templates and visual workflows cover no-code assembly; deeper integration needs SDK calls and CLI configuration.

Are third-party models available?

Yes. Alongside in-house Qwen and Wan models, the platform includes DeepSeek, Kimi-K3, GLM-5.2 and others.

How do I control cost?

Send deferrable work to batch, enable caching for repeated prompts, and size reserved throughput to real peaks instead of over-provisioning.