Enterprises do not want a model. They want a workflow that answers the phone, checks the form, and flags the bad batch—without hiring a research team to wire it together.
Baidu Smart Cloud's answer is Qianfan (千帆), its enterprise large-model (大模型) platform, upgraded to 4.0 in late August 2025.
The shift in this release is small in wording but large in intent: Qianfan now describes itself as agent-first.
Key takeaways
- Qianfan (千帆) is Baidu Smart Cloud's enterprise platform for building and running large models (大模型); version 4.0 was unveiled at the 2025 Baidu Cloud Intelligence Conference in Beijing on 28 August 2025.
- Baidu positioned 4.0 as an agent-centric, one-stop enterprise service platform covering models, agent orchestration, data, and enterprise-grade controls.
- The model library offers more than 150 models, including Baidu's self-developed video-generation model ("Steam Engine," 百度蒸汽机) and specialist models for finance and vision.
- Baidu reported over 460,000 enterprise users and more than 1.3 million agents built on the platform.
- Agent tooling includes multimodal RAG, a multi-agent collaboration framework, and MCP services such as Baidu's AI Search MCP Server.
- A new RFT (reinforcement-feedback fine-tuning) toolchain is said to cut fine-tuning data needs from thousands of examples to hundreds.
From model marketplace to agent factory
Earlier enterprise AI platforms mostly sold access to models and left the assembly to the customer. Qianfan 4.0's pitch is that the agent (智能体) is the product: developers get model access, orchestration, data hooks, and governance in one place. That matches where enterprise demand has actually gone—most companies do not want to tune a foundation model, they want a dependable component that performs a job.
Why enterprises buy platforms, not models
The interesting competitive question is not whether Qianfan has the single best base model—Baidu openly hosts rival models alongside its own ERNIE. The value proposition is the plumbing: identity, permissions, audit logs, and the ability to swap the underlying model without rewriting the application. For a bank or a manufacturer, a platform that satisfies compliance and keeps data inside its own perimeter is worth more than a slightly better benchmark score. That is why Baidu leads its pitch with agent counts and model breadth rather than one headline model. It also explains the RFT toolchain's emphasis on small data: the blocker for most enterprises is not model quality but having enough labelled examples, so lowering that bar is a direct attack on the real adoption wall. It is a bet that the winner in enterprise AI is the one who owns the workflow, not the weights.
The numbers behind the claim
The scale figures are the ones worth scrutiny. Baidu says the platform serves over 460,000 enterprise users and has seen more than 1.3 million agents created. A model library of 150+ covers deep reasoning, vision understanding, vision generation, and speech. These are company-reported totals from the launch (carried by Xinhua and China Securities Journal), so read them as platform cumulative counts rather than active production deployments.
What "agent-centric" gives a builder
Concretely, 4.0 adds multimodal RAG (retrieval-augmented generation that can pull from text and images), a multi-agent collaboration framework for slower "think hard" tasks, and MCP services—including Baidu's own AI Search MCP Server—so an agent (智能体) can call external tools through a standard interface. The multi-agent framework matters most for slow, multi-step jobs—research briefs, reconciling records—where a single model call is not enough and a coordinator has to plan, delegate, and check the result. There is also a fine-tuning method, RFT, that Baidu says reduces the data required for specialization from thousands of labelled examples to hundreds, lowering the bar for a company-specific model.
A reasoning agent on top
Separately, Baidu's "Agent Pro" (智能体Pro), a reasoning agent first shown at the Create 2025 developer conference in April 2025, was later upgraded to plug into ERNIE 4.5 with 100+ prebuilt components and a low-code framework. It is the customer-facing face of the same agent strategy: configure, don't code from scratch.
Honest limitations
The platform scale and feature claims come from Baidu's launch at the 2025 Cloud Intelligence Conference, reported by Xinhua, China Securities Journal, and Science and Technology Daily. The 460K-enterprise and 1.3M-agent figures are Baidu's own cumulative counts; we did not audit them, and "agents built" is not the same as "agents in production." The competitive framing (Baidu cites IDC placing its AI public cloud at a 24.6% share, six years running) is a third-party claim repeated in coverage, not independently verified here. Capabilities move fast across platform versions.
What readers can do now
- If you run an enterprise AI pilot, evaluate a managed agent platform like Qianfan against building orchestration in-house—the 150+ model library and MCP hooks remove a lot of plumbing.
- Before committing, ask vendors to separate "agents created" from "agents in production"; the two numbers tell very different stories.
- Use RFT-style fine-tuning for a narrow internal task only if you have even a few hundred good examples; otherwise start with retrieval over your own data.