ModelScope is a Chinese open-source model community where developers download models and datasets, host Studios demo apps, and call selected models through an API metered in free daily Magicubes.
Used this tool? Rate it
Used this tool? Rate it
ModelScope (魔搭社区) is an open-source model hub and community where developers find, download and publish AI models and datasets, run hosted demo apps called Studios, and call selected models through an API.
ModelScope says it was founded by the Institute for Intelligent Computing in June 2022, together with the CCF Open Source Development Committee, as an open-source model community and innovation platform. The contracting party in the user agreement is Shanghai ModelScope Technology Co., Ltd.
The main site is modelscope.cn. Alibaba Cloud has opened ModelScope to English-language users; before that, trade press reported, the platform had been limited to users in mainland China. A separate international site at modelscope.ai runs its own documentation set, and its API-Inference rules repeat the Alibaba Cloud account requirement described below for modelscope.cn.
Three things shape most decisions: free use is metered in Magicubes, a compute credit issued for logins and community activity; the free API and free Notebook GPU both depend on a linked Alibaba Cloud account; and each model carries its own license rather than one platform-wide license.
The company describes the hub as a place to explore the industry's latest models, datasets and AI applications and to access computation infrastructure for model serving. Models can be public or private: a model created through the command-line upload tool is public by default, and visibility can be switched to private afterwards on the model's editing page.
Publishers can set custom license terms that bind anyone who downloads or uses their files. When a repository is created through the Python SDK without naming a license, the SDK falls back to Apache-2.0.
The ModelScope API service, API-Inference, serves selected open-source models behind standardized endpoints so they can be tried without a local GPU and wired into applications. Its coverage includes selected large language models, multimodal models and text-to-image models from the AIGC zone.
Supported models carry a "blue-green lightning" API-Inference badge, and the sample code on each model page is the authoritative calling reference, since reasoning models can differ slightly from standard LLM calls.
Studios are hosted spaces for showing AI applications built on ModelScope models. Pages are built with the platform's visualization SDKs, currently Gradio, Streamlit and a Static HTML SDK, so one Studio can combine several models or expose many parameters. New Studios do not enter the featured list on the Studios homepage automatically; inclusion is by application.
ModelScope Notebook is a cloud development environment based on Jupyter Notebook, run on cloud CPU or GPU instances in partnership with Alibaba Cloud PAI-DSW. It comes with ModelScope's development packages and algorithm libraries pre-installed, allows extra third-party libraries, and adds a Web-IDE entry for online development in VS Code.
The AIGC zone covers image generation (Agentic, quick and professional modes) and video generation, plus training for image models, image-editing models and video models. Both inference and training there are paid in Magicubes. The same reward rules also count MCP services, Skills and Studios as community contributions.
The open-source ModelScope Python library offers one interface for inference, fine-tuning and evaluation across computer vision, NLP, speech, multimodal and scientific-computing models. Swift is the community's LLM training and inference framework, supporting model families such as LLaMA, Qwen, ChatGLM and Baichuan and methods such as LoRA; EvalScope is a separate framework for model evaluation and performance benchmarking.
ModelScope's use cases map onto its separate services, each with its own quota and limits:
It fits less well for:
On modelscope.cn, API-Inference is provided free of charge to registered users, and its calls, like AIGC-zone generation, are paid in Magicubes. Magicubes come in two kinds: short-term Magicubes last 24 hours and long-term Magicubes last 90 days. Using them first requires a completed profile with a verified email address.
| How Magicubes are earned | Amount | Type and cap |
|---|---|---|
| Daily login | 200 per day | Short-term, valid that day, once per day |
| Linked Alibaba Cloud account | 50 per day | Short-term, valid that day, once per day |
| Verify email | 50 | Long-term, one time |
| Popular model (100+ downloads and 5+ likes) | 50 | Long-term, no cap |
API-Inference deducts Magicubes per call in three tiers set by model size and compute needs.
| API-Inference model tier | Average cost per call |
|---|---|
| Lightweight models | 0.5 Magicubes |
| Mainstream models | 1 Magicube |
| Flagship models | 2 Magicubes |
With a sufficient Magicube balance, API-Inference can be called any number of times. Both kinds of Magicube count the same when spent, and the ones closest to expiry are deducted first. AIGC-zone costs scale with model size, image count, video resolution and length, or training steps. Premium options multiply the cost; priority training, for example, costs 4× Magicubes.
If a task fails because of a system error, the full Magicube cost is refunded. A task the user cancels is treated differently: cancelled inference is refunded, but a cancelled training task still consumes its Magicubes.
Notebook uses a separate quota. Each new user gets a PAI-DSW CPU environment with 8 cores and 32 GB for long-term use, plus a GPU environment with 8 cores, 32 GB RAM and 24 GB of GPU memory limited to 36 free hours. Once that is used up, the commercial Notebook edition is bought in the relevant Alibaba Cloud product console; the Notebook page does not list its price. Studios are presented as a free space for showcasing AI applications.
For high concurrency or large quotas, the API-Provider option binds an external API provider instead of relying on ModelScope's free service.
Independent coverage names Hugging Face and Amazon Bedrock as the closest comparisons, noting that ModelScope's pitch rests on developer tooling plus Chinese-language datasets and data processing toolkits.
| Option | Scale or focus | Overlap and access |
|---|---|---|
| ModelScope | More than 5,000 models from Alibaba and other Chinese AI firms at its English launch, trained largely on Chinese | Free API calls metered in Magicubes, with a real-name Alibaba Cloud account |
| Hugging Face | Over 2M models, 1.5M datasets and 1.5M AI apps (Spaces), all open and publicly available | Some ModelScope models are also published there |
For anyone weighing ModelScope as a Hugging Face alternative, the split is catalog focus and account setup: ModelScope concentrates models from Chinese labs and ties its free API to Alibaba Cloud.
Largely. API-Inference is free for registered users, but each call deducts Magicubes, which are issued for free: 200 short-term Magicubes each day on login and 50 more once an Alibaba Cloud account is linked. New users also get 36 free hours on a 24 GB GPU Notebook; beyond that, Notebook GPU time is bought through Alibaba Cloud.
For the free inference API and the free Notebook quota, yes. API-Inference additionally requires that the linked cloud account has passed real-name verification.
It depends on each model's own license, which users agree to follow; some are permissive and some non-commercial. API-Inference itself is described as a non-commercial service.
Yes, for large language models: the ModelScope access token acts as the API key and a ModelScope Model ID as the model name. An Anthropic-compatible mode is in beta.
