Meet MOSS: China’s Homegrown ChatGPT Rival and Its Capabilities

ChatGPT is the most advanced AI and the hottest application, reaching over 100 million monthly active users within two months of its November‑2022 launch, setting a record for user growth speed on the internet.

Because training such models requires massive compute and costly annotation, domestic equivalents have not yet been released to the public, though major Chinese internet companies have announced plans to build “domestic ChatGPT” products.

On the night of February 20, the Natural Language Processing Laboratory at Fudan University announced MOSS, a language model with ChatGPT‑level abilities, and opened it for public beta testing.

MOSS experience link: https://moss.fastnlp.top/

MOSS project homepage: https://txsun1997.github.io/blogs/moss.html

The name MOSS comes from the movie “The Wandering Earth,” and the announcement quickly topped Zhihu’s hot search list.

Due to overwhelming demand, the MOSS servers were overloaded shortly after launch, showing the high public expectation for generative language models.

According to Fudan researchers, the current beta is still in internal testing and not yet ready for a public release.

Core Functions

MOSS provides capabilities similar to ChatGPT, handling tasks such as text generation, summarization, translation, code generation, and casual conversation. Access is free during the preview period.

Model Architecture

The development process includes two stages: training a foundational language model and then training dialogue abilities to understand human intent.

Key Differences from ChatGPT

MOSS has far fewer parameters than ChatGPT.

It learns by interacting with humans and other AI models, whereas ChatGPT relies on Reinforcement Learning from Human Feedback (RLHF).

MOSS will be open‑source to promote future research, while ChatGPT may remain closed.

Interaction Examples

In one example, a user asked MOSS to recommend five sci‑fi movies, generate a table with directors, and then insert a column for release years, demonstrating strong multi‑turn interaction and instruction understanding.

MOSS sometimes produces factual errors, such as misidentifying the director of “The Matrix” as Thomas Neff instead of the Wachowski siblings.

For code generation, MOSS can provide a Python implementation of quicksort and explain the code upon request, effectively acting as a hands‑on programming tutor.

It can also answer detailed questions about provided code, correctly identifying the programming language and function purposes.

When faced with unreasonable requests, MOSS refuses to comply and offers appropriate guidance, reflecting alignment with human values.

Training Details

MOSS is trained with a self‑developed model containing on the order of tens of billions of parameters. During dialogue training, OpenAI collected hundreds of thousands of human instructions for ChatGPT; the Fudan team instead let MOSS interact with humans and other AI models, significantly improving learning efficiency and development speed.

Current Limitations

Limited multilingual corpora cause weaker performance in non‑English languages, especially Chinese.

Smaller model capacity leads to insufficient world knowledge, resulting in occasional misleading or false responses.

The model sometimes fails to follow instructions precisely, requiring users to regenerate or re‑prompt.

Occasional generation of unethical or harmful content can occur; users can provide feedback to mitigate this.

The research team notes that MOSS’s English output is stronger than its Chinese output because the base model was trained on over 300 billion English tokens versus about 30 billion Chinese tokens.

Future Directions

Future work aims to add multimodal capabilities such as drawing, speech, and music generation, and to enhance assistance for scientific research.

Team Introduction

The main authors are Professor Qiu Xipeng and his Ph.D. student Sun Tianxiang from Fudan University, along with several other contributors.

Professor Qiu Xipeng is a distinguished researcher in natural language processing and deep learning, with numerous high‑impact publications and awards.

Sun Tianxiang focuses on efficient fine‑tuning and inference of pretrained language models, and has published papers at top conferences such as ICML, ACL, and AAAI.