01.AI launches the Wanzhi enterprise LLM platform

What was announced

On 17 March 2025, 01.AI launched the Wanzhi enterprise one-stop large-model platform, together with enterprise-grade DeepSeek deployment solutions ^[1]. Xinhua reported the public release of the platform that day, and that it also included customized DeepSeek deployment for enterprise customers ^[1]. One-stop here means packaging the pieces an enterprise needs — model deployment, application practice and fine-tuning tools — into a single offering, instead of leaving each customer to assemble them separately.

What the platform contains

Wanzhi is not built around a single model. Alongside DeepSeek, it supports Alibaba's Qwen and 01.AI's own Yi models ^[1]. The published scope of the solution covers DeepSeek model deployment, application practice and model fine-tuning tools ^[2].

The three parts correspond to different stages of enterprise use. Deployment is about getting a model running in an environment the customer controls. Application practice is about turning a general-purpose model into something that fits a concrete workflow. Fine-tuning is the step in which a model already trained on large amounts of general text is adjusted further on a company's own data, so that it behaves better on a narrow task; it is a way to obtain specialized behaviour without training a model from scratch.

Supporting several model families on one platform changes the position of the platform vendor. The customer is not locked into a single model, and the platform's job becomes managing deployment and adaptation across whichever models the customer picks ^[1].

The business and model strategy behind it

Kai-Fu Lee said that 01.AI's ToB (enterprise-facing) revenue had exceeded 100 million yuan in the previous year, and that first-quarter revenue was close to the whole of the previous year ^[3]. The company also said it would not build trillion-parameter models, and would focus on lightweight models instead ^[3].

The two statements are connected. A lightweight model is cheaper to run and easier to deploy and adapt than a very large one, which suits the kind of on-premise, task-specific work that enterprise customers ask for. A company that opts out of the race to build the largest models is in effect putting its effort into models that can be delivered and customized at a workable cost, and into the tooling around them. The enterprise platform is the commercial expression of that choice: the product being sold is not one flagship model but the ability to put models — including third-party ones — to work inside a customer's systems ^[1][2].

Why it matters

For an enterprise, the practical question is rarely which model tops a leaderboard. It is which model can be deployed, adapted to internal data and maintained over time without rebuilding everything when a new model appears. A platform that offers DeepSeek alongside Qwen and Yi, with deployment, application practice and fine-tuning tools attached, answers that question by making the model a replaceable component rather than a fixed commitment ^[1][2].

It also marks a division of labour in the market. Model developers supply the models; platform providers supply the deployment, adaptation and integration work that turns a model into a working system. 01.AI's decision to concentrate on lightweight models rather than trillion-parameter ones points in the same direction: the company is competing on what can be delivered to customers, not on the largest possible training run ^[3].