ModelScope

What ModelScope is

ModelScope (魔搭) is an open-source model community and innovation platform. Its own site dates its founding to June 2022, when the Institute for Intelligent Computing set it up together with the CCF Open Source Development Committee^[2]. What the platform collects is not code projects but models: trained, validated models that a developer can pick up and run, along with the pages, examples and tooling needed to try them out.

How it started

ModelScope was officially released at the 2022 Apsara Conference on 3 November 2022. Alibaba's DAMO Academy put it out together with the CCF Open Source Development Committee and partners including 澜舟科技, 深势科技, 智谱AI and 中国科学技术大学. The announcement was made by 周靖人, Alibaba senior vice president and DAMO Academy vice president^[1].

The underlying code repository, modelscope/modelscope on GitHub, was created earlier, on 25 July 2022, and is released under the Apache-2.0 license^[3]. Interest showed up before the formal launch: 澎湃新闻 reports that a small developer evaluation opened in August 2022 and drew 400,000 to 500,000 page views in under three weeks^[1].

What it offers

At launch, DAMO Academy had open-sourced more than 300 validated models on ModelScope. More than 150 of them were state-of-the-art models, and more than a third were Chinese-language models^[1]. The emphasis on Chinese-language models reflects a practical gap: much of the published model work at the time was built and documented for English, so a community where Chinese-language models are a large share of the catalogue serves developers working in that language directly.

The platform is also built to lower the cost of getting a model to actually run. It offers one-click model experience, inference in one line of code and fine-tuning in ten lines of code^[1]. In practice that means a user can open a model page and try it without setting up an environment, call a hosted or packaged model from a short script, and adapt a pretrained model to their own data without writing a training pipeline from scratch.

Model as a service

The platform is tasked with bringing the concept of Model-as-a-Service to life and building a prosperous AI ecosystem around open-source models^[2]. The idea behind that phrase is a shift in how models are consumed. Instead of every team training and maintaining its own model, models are treated more like a service: someone publishes and maintains them, and others call them, adapt them, or build on top of them. That only works if the models are easy to find, easy to try and easy to modify, which is what the one-click and few-line-of-code features are aimed at.

Where it sits

ModelScope is an effort by a corporate research lab to open its model collection to outside developers, rather than keeping the models inside its own products. It brings together a research institute, a professional society's open-source committee, several smaller AI companies and a university as launch partners^[1]. For developers, the practical value is a single place where a few hundred validated models, a substantial share of them in Chinese, can be tried and fine-tuned; for the wider field, it is one of the ways trained models move from papers and internal repositories into public, reusable infrastructure.