ModelScope
Alibaba's open model hub (魔搭社区). Apache 2.0 client library, 170,000+ open-weight models and 2,000 free API calls a day. Where most Chinese labs publish first.
Overview
Alibaba DAMO Academy built ModelScope as a model-as-a-service hub in 2022 and it has since become the default publishing venue for Chinese open-weight releases. The numbers reported in April 2026 put it past 170,000 models and 25 million registered developers. For an engineer the important part is not the count. It is that Qwen, DeepSeek, GLM, MiniMax and a long list of smaller labs treat ModelScope as a first-class target rather than a mirror, so weights, quantisations and community fine-tunes frequently appear here before anywhere else.
The tooling around it is the underrated half. The modelscope client is Apache 2.0 and deliberately shaped like huggingface_hub, so snapshot_download and from_pretrained calls transfer with minimal edits. ms-swift, the sibling training framework, is also Apache 2.0 and covers 600+ text models and 300+ multimodal ones with LoRA, QLoRA, DoRA, full-parameter tuning, DPO and GRPO. Both ship maintained Ascend NPU images, which makes this one of the few complete fine-tuning paths that does not assume Nvidia hardware. The v1.40.0 release on 28 August 2026 included an Ascend image rebuild and Docker support for ms-swift 4.4.x, which tells you where the maintenance attention is going.
Two practicalities decide whether this fits your team. First, registration binds to an Alibaba Cloud account with real-name verification, so a foreign team without a mainland phone number or a willingness to upload a passport will hit friction on day one. Second, the Apache 2.0 licence covers the code, not the weights. A meaningful fraction of hosted models carry bespoke community licences with monthly-active-user caps or restrictions on specific applications, and the model card is the only authoritative source. Treat the hub as infrastructure and the licence review as a separate, per-model task.
Key Features
- ✓ Apache 2.0 Python client at v1.40.0, released 28 August 2026, with a snapshot_download API close enough to the Hugging Face Hub that most scripts port on a single import line
- ✓ 170,000+ open-weight models as of April 2026 across NLP, vision, speech and multimodal, weighted heavily toward Chinese-origin releases
- ✓ API-Inference free tier of 2,000 calls per day spanning 900+ hosted models including Qwen, DeepSeek, GLM and MiniMax
- ✓ ms-swift, the Apache 2.0 companion training framework, covering 600+ LLMs and 300+ multimodal models with LoRA, QLoRA, DoRA, full fine-tune, DPO and GRPO
- ✓ Ascend NPU Docker images built and maintained in the repository, so fine-tuning has a real non-CUDA path
- ✓ Free GPU studio hosting for Gradio, Streamlit and Docker apps with no code changes, plus dataset hosting
Where it holds
- • A large share of Chinese labs push to ModelScope on release day and mirror to Hugging Face later, sometimes days later and occasionally not at all
- • Inside mainland China the hub pulls at full line speed while Hugging Face needs a tunnel, which changes your CI story completely
- • The client library and ms-swift are actually Apache 2.0, not source-available with commercial strings attached
- • Ascend NPU support is maintained in-tree rather than living in a community fork
Where it breaks
- • Registration binds to an Alibaba Cloud account with real-name verification, which in practice means a mainland phone number or a passport upload
- • English documentation trails the Chinese pages by weeks and several SDK guides have no English version at all
- • Platform licence and weight licence are different things: plenty of hosted models ship custom community licences with commercial caps or field-of-use limits, so every model card needs reading
- • Quality is uneven across a very long tail of stale or abandoned uploads, and there is no strong signal separating maintained repos from dead ones
My Take
Proximity is what you are actually buying. From inside China this hub pulls at full speed while Hugging Face needs a tunnel, and by April 2026 it held over 170,000 models, a lot of which appear here first. The client library is Apache 2.0, v1.40.0 landed on 28 August 2026, and the download API is close enough to Hugging Face that porting a script is usually one import line. Budget an afternoon for the real-name registration, then read every model card, because the platform licence and the weight licence are not the same document.
Quick Info
- Pricing:
- freemium
- Openness:
- Open source
- Licence:
- Apache 2.0 (client library and ms-swift; hosted model weights carry their own separate licences)
- Starting at:
- Model and dataset hosting is free with unlimited storage. API-Inference gives 2,000 free calls per day across 900+ hosted models with no card required, then moves to paid quota. Free GPU-backed app hosting for Gradio, Streamlit and Docker studios, plus free LoRA training runs for image and video models. Paid Alibaba Cloud compute (PAI, EAS) sits behind the one-click deploy buttons and is billed on the normal Alibaba Cloud meter. The Python library and ms-swift are free forever under Apache 2.0.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where ModelScope holds and where it breaks. The fastest check is your own workload.
Visit website ↗