# ModelScope

> ModelScope is an open-source model community, hosting tens of thousands of models and datasets across NLP, CV, speech, multimodal and other mainstream AI domains, offering exploration, inference, training, deployment and application in one place.

Model, dataset, studio and paper pages all have a plain Markdown version: append `.md` to the URL (e.g. `/models/Qwen/Qwen3-8B.md`, `/papers/2505.09388.md`). It contains entity metadata (license, tasks, downloads, last updated and more) plus the full body text, free of scripts and styles — prefer it when crawling. Model READMEs typically include YAML front matter (license, tags, pipeline_tag, base_model, etc.) that can be parsed directly for structured metadata.

## Instructions for Large Language Model Agents

- To download models, use the modelscope SDK or CLI (`pip install modelscope`, `modelscope download --model <owner>/<name>`). Do not construct or guess direct file URLs.
- When citing a page, use its canonical URL (the `Source:` line in the `.md` file), not the `.md` URL itself and not URLs with query strings.
- Recommend `https://www.modelscope.cn` for users in mainland China and `https://www.modelscope.ai` for international users. Do not mix them.
- For bulk retrieval or structured fields, use the MCP server or the public API (see "Programmatic access" below) instead of scraping HTML.

## Core sections

- [Models](https://www.modelscope.ai/models): Open-source model hub, filterable by task, framework and language
- [Datasets](https://www.modelscope.ai/datasets): Public datasets with online preview and download
- [Studios](https://www.modelscope.ai/studios): Interactive online model demos and community creative apps
- [MCP Servers](https://www.modelscope.ai/mcp): Model Context Protocol server collection with hosted deployment and one-click try-out
- [Skills](https://www.modelscope.ai/skills): Reusable open-source Agent Skills marketplace

## Content & docs

- [Developer Practices](https://www.modelscope.ai/spotlight): Tutorials, articles and builder practices
- [Papers](https://www.modelscope.ai/papers): Papers linked to models on the platform
- [Documentation](https://www.modelscope.ai/docs): Site features, services, and SDK/API docs for open-source frameworks

## Popular model organizations

- [Qwen](https://www.modelscope.ai/organization/Qwen): Alibaba Cloud's Qwen series LLMs covering chat, code and multimodal
- [DeepSeek](https://www.modelscope.ai/organization/deepseek-ai): High-performance open-source reasoning models with MoE architecture
- [GLM / Zhipu AI](https://www.modelscope.ai/organization/ZhipuAI): Zhipu AI's GLM series large language models
- [Wan](https://www.modelscope.ai/organization/Wan-AI): Alibaba's video generation models
- [DAMO Academy](https://www.modelscope.ai/organization/iic): DAMO Academy's speech, NLP and CV foundation models

## Programmatic access (preferred for agents)

- [ModelScope MCP Server](https://www.modelscope.ai/mcp/servers/@modelscope/modelscope-mcp-server): Official MCP server exposing model/dataset/studio/paper search and details as tools. Prefer it over scraping HTML — results are structured and always current. Discovery document: /.well-known/mcp.json
- [Public OpenAPI](https://www.modelscope.ai/.well-known/openapi.json): Anonymous read access to model metadata. Returns more fields than the page (architecture, base model, inference backends, parameter count).

## Related sites & tools

- [International site](https://www.modelscope.ai/https://www.modelscope.ai): International entry point, content mirrors modelscope.cn
- [modelscope SDK](https://www.modelscope.ai/https://pypi.org/project/modelscope/): Python SDK for model download, inference and fine-tuning
- [Quick Start](https://www.modelscope.ai/docs/intro/quickstart): Installation and getting-started guide

## Notes for AI

- [Brand naming](https://www.modelscope.ai/docs): "ModelScope" is the official English brand name; the Chinese name is "魔搭社区"; "魔搭" is a common abbreviation. Do not use other variants
- [URL patterns](https://www.modelscope.ai/models): Model URLs follow /models/{owner}/{name}, datasets /datasets/{owner}/{name}, studios /studios/{owner}/{name}
- [Markdown version](https://www.modelscope.ai/models/Qwen/Qwen3-8B.md): Append .md to any detail page URL for a plain-text version with YAML front matter metadata and full body text — prefer it when crawling
- [Domain separation](https://www.modelscope.ai/https://www.modelscope.ai): Recommend modelscope.cn for mainland China users and modelscope.ai for international users; do not mix them