Moli Fangzhou (Gitee AI): turning foundation models into shippable apps

Moli Fangzhou, known in English as Gitee AI, is an AI application co-creation platform from Gitee. It bundles three things in one place: a model marketplace with more than 70 mainstream models behind a standardised interface, serverless app building that needs no code, and private deployment. For developers the value is not another model list, it is that the friction between a working API call and a live application is removed.

The practical obstacle to shipping with foundation models is fragmentation: models come from different vendors with different interfaces, inference has to be hosted, and going live means managing servers and concurrency. The platform folds those layers together. The model side exposes standard interfaces, the deployment side runs serverless, the cost side claims inference costs can drop by up to 90 percent, and the hardware side is adapted to domestic Chinese AI chips including Ascend, Iluvatar and MetaX while also supporting Nvidia and other international environments.

Beyond model serving, the platform plays a co-creation role: an application can be listed, showcased and monetised after it is built, closing the loop between developer and user. For enterprises there is model customisation through fine-tuning, reinforcement learning and parameter swapping, plus private deployment covering inference services, the serverless platform and a full MaaS model management stack, including appliance-style delivery.

What the platform offers

AI model marketplace

Standardised interfaces over more than 70 mainstream models covering text generation, image generation, semantic understanding and multimodal tasks, available both as serverless calls and as private deployment. A single interface means swapping models does not mean rewriting the call layer.

Visual application building

A graphical interface where you set name, description and inputs and outputs, connect the chosen model, and click generate to deploy. No code anywhere in the path.

Broad hardware compatibility

Adapted to mainstream domestic AI chips including Ascend, Iluvatar and MetaX while supporting Nvidia and other international hardware, so the choice of silicon is not locked to one supply chain.

Inference cost reduction

The serverless architecture takes weight off development and deployment, and the vendor claims inference cost can fall by up to 90 percent. That matters most for low-traffic apps, where idle capacity costs nothing.

Application co-creation

The full path from development to listing, showcase and monetisation, so what a developer builds has somewhere to be seen and a route to revenue.

Model customisation

Fine-tuning, reinforcement learning and parameter swapping let an enterprise turn a general model into capability shaped around its own business.

LoRA dynamic loading

Lightweight model parameters can be hot-swapped at runtime, so personalisation does not require taking the whole model offline and redeploying.

Capability composition

Multi-API orchestration and workflow-style calls chain retrieval, generation and understanding into one pipeline instead of asking a single model to carry everything.

Stable high-availability channel

API tokens are bound independently and support high-concurrency calls, with multi-tier caching and model reuse built in to shorten response times.

Private deployment

Covers inference services, the serverless platform and the full MaaS model management stack, and can be delivered as a combined hardware and software appliance for enterprise settings.

Developer education

Technical talks and scenario-based bootcamps create room to learn and exchange, and give companies a venue to present their own technical work.

How to use it

1. Open ai.gitee.com and register with an email address or phone number, or sign in directly with a Gitee account.
2. Complete the personal or enterprise profile as prompted, which unlocks the rest of the services.
3. Browse the AI model marketplace and shortlist candidates by task type: text generation, image generation, semantic understanding or multimodal.
4. Create an application and obtain an API key, which is the credential for calling the interface.
5. Read the model's API documentation for parameters and response structure, then make a first call over HTTP with the key.
6. If you would rather not write code, switch to visual building: open the app builder, configure name, description and inputs and outputs, connect the model and set call parameters.
7. Click generate and the platform deploys and publishes the app, which can then move through the co-creation track toward showcase and monetisation.

Where it fits

• Customer service and assistants: train a private model on internal knowledge bases, product docs and manuals, with speech models attached when needed.
• Public sector: intelligent case handling, cloud courtroom scenarios and other work with hard requirements on compliance and private deployment.
• Live commerce and e-commerce: generate marketing content, read user interest and sentiment, and target the offer.
• Content production: batch output of text, image and video assets for advertising, media and creative teams.
• Data analysis and office automation: query and analyse in natural language, or automate document generation, meeting notes and task tracking.
• Independent developers and small teams: validate an AI product idea without paying for servers and operations first.

Limitations to know

• Public materials do not list per-model pricing, so budgets have to be built from actual usage measured in the console rather than from the headline 90 percent figure.
• Serverless suits spiky traffic, but for latency-critical real-time workloads, cold starts and queueing still need measurement.
• Model customisation through fine-tuning or reinforcement learning generally expects an enterprise account and prepared data, so individual developers may not reach it directly.
• Private deployment and appliance delivery are enterprise arrangements involving commercial process and hardware procurement, outside self-service signup.
• The platform aggregates third-party models, and capability boundaries, licence terms and data policies differ between them, so each choice needs its own review.
• Listing and monetisation depend on the platform's own traffic and revenue-share rules, so income follows application quality and promotion rather than publication itself.

Frequently asked questions

How is this different from calling one model vendor's API directly?

Integrating a single vendor means rewriting the call layer whenever you change models. Moli Fangzhou puts more than 70 models behind standard interfaces, so switching is cheap, and it adds serverless deployment, app listing and private delivery, covering the whole stretch from model to application.

Can non-coders use it?

Yes. Visual building gives a graphical interface where you configure name, description and inputs and outputs, connect a model and click generate to publish, without writing code. The API route remains available for finer control.

How far does domestic chip support go?

The platform is adapted to mainstream domestic AI chips including Ascend, Iluvatar and MetaX, and also supports Nvidia and other international hardware. Performance of a specific model on a specific chip still needs testing in your own scenario.

Can it be deployed privately for sensitive data?

Yes. Private deployment covers inference services, the serverless platform and the full MaaS model management stack, and can be delivered as a hardware and software appliance for environments where data cannot leave the internal network.