ModelScope

What is the ModelScope?

ModelScope is a joint venture between Aliyun and the Open Source Development Committee of the Chinese Computer Society to develop AILarge model open sourceSince its official release at the 2022 Yunqi Conference, MaaS has been practicing the concept of "Model as a Service (MaaS)" and is committed to lowering the threshold of AI applications and promoting the construction of a large model ecosystem. Its core goal is to provide developers with one-stop services such as model experience, download, tuning, training, reasoning, deployment, etc. by means of open source and community building, so as to make AI technology accessible.

History and Milestones

• November 2022The first batch of over 300 verified and high-quality AI models, of which Chinese models account for more than one-third, covering more than 60 tasks, such as vision, speech, natural language processing, and multimodal, are on the official launch of the ModelScope.
• 2023The number of models in the community has exceeded 2,300, the size of developers has exceeded 2.8 million, and the number of model downloads has exceeded 100 million. The industry's first open-source model for text to video generation, Baichuan Intelligence, and the Shusheng series of large models were all premiered in the community.
• 2024The "ModelScope-Sora Open Source Program" was launched to help explore Sora-like models in China; the AIGC zone was released to provide image and video creation and model training services; and the Step-Video video generation and Step-Audio voice model adaptation was completed with Huawei Rise.
• 2025The total number of models in the community exceeds 70,000, serving over 16 million developers in 36 countries; 4,000+ services are online in the MCP Plaza, and the call volume exceeds 100 million times; DeepSeek series of models, Step Star multimodal model, and so on, are premiered in the community.

Core functions and services

1. Model repository

• Areas covered: natural language processing (NLP), computer vision (CV), speech recognition, multimodality, scientific computing, etc.
• Model Type::
  • Pre-trained models: e.g. Tongyi Qianqi, Beanbag Model, Kimi, etc., covering billions to tens of billions of parameter sizes.
  • open source model: Support developers to share research results, such as UniASR speech recognition and SAMBERT speech synthesis.
  • Chinese model: Over 100 Chinese large models, including Chinese GPT-3, Chinese version of CLIP, etc. to meet localization needs.
• typical model::
  • NLP: StructBERT, mPLUG, Chinese GPT-3, and others.
  • CV: DCT-Net portrait cartooning, image description generation model.
  • colloquial (rather than literary) pronunciation of a Chinese character: UniASR, Paraformer non-autoregressive speech recognition models.
  • multimodal: OFA, Team, mPLUG, etc. graphical understanding and generation of models.

2. Data sets and tool chains

• data set: Provide massive labeled data, covering text, image, audio, video and other multimodal, to support model training and validation.
• toolchain::
  • Python SDK: Encapsulate efficient inference, fine-tuning and evaluation functions, one line of code to realize model inference, ten lines of code to complete the tuning task.
  • Pipeline Mechanism: Quickly build complex tasks such as text categorization and sentiment analysis through simple API calls.
  • Cloud Notebook: Provide a low-cost development environment that supports online model training and debugging.

3. Deployment and reasoning services

• Cloud Deployment: Supports one-click deployment of models to AliCloud to generate API interfaces for application calls.
• Seamless integration between local and cloud: Supports both quick installs for local environments and cloud-based Notebook installation-free development.
• Elastic arithmetic support: Relying on AliCloud EAIS and PAI platforms, it provides efficient training and inference services.

4. Community collaboration and ecology

• open source community: Developers can share models, datasets, participate in community governance, and drive technology iteration.
• Cooperation between universities and colleges: Jointly develop courses and support academic research with more than 20 universities, including Tsinghua University and Peking University.
• Events & Challenges: Regular competitions on model optimization, application development, etc. are held to stimulate innovation.
• MCP Plaza: Aggregate thousands of MCP services and hosting solutions, support third-party integration, such as Alipay, MiniMax and other innovative services exclusive debut.

Technological Advantages and Innovations

1. Accessibility

• One-Stop Service: Integrate model exploration, environment installation, inference validation, training tuning, etc., and shorten the development cycle from days to hours.
• Low-code development: Quickly build AI applications by dragging and dropping components, combining models, etc. without writing complex code.
• Visualization tools: Provide model cards, online experience windows, etc., to lower the threshold of technical understanding.

2. Performance leadership

• arithmetic support: Relying on AliCloud Elastic Accelerated Computing (EAIS) and Machine Learning Platform (PAI), it provides efficient training and inference services.
• model optimization: Reduce the cost of fine-tuning and improve model adaptability through techniques such as LoRA and ResTuning.
• multimodal support: Realize cross-modal combination of text, image, and speech to support complex scene applications.

3. openness and compatibility

• framework neutrality: Compatible with TensorFlow, PyTorch, Caffe and other mainstream frameworks, supporting multi-model training and service deployment.
• Data transparency: Make the model download data sources and calculations publicly available for community scrutiny.
• Global collaborationThe company has attracted the participation of overseas developers and become the domestic version of "HuggingFace".

User Feedback and Industry Impact

1. Developer evaluation

• powerful: Chinese models are rich and suitable for localization needs.
• Convenient operation: Documentation and examples are friendly and newbies get started quickly.
• Active community: Issues are responsive and collaborative.

2. Industry recognition

• Model Performance: SOTA results in tasks such as VQA and machine translation.
• Ecological Impact: Promote AI technology in healthcare, education, industry and other fields, such as protein model-based drug development.
• international influence: Serving developers in 36 countries and becoming the world's leading open source ecological platform for AI.

Through the concept of "Model as a Service", ModelScope continues to lower the threshold of AI technology and promote the application of large models in thousands of industries. With the launch of MCP Square, the community has further expanded its model componentization capabilities to support developers in building more complex AI systems. In the future, MCP is expected to become the platform of choice for global AI developers and promote the universal development of AI technology.