MOSS is an open-source bilingual conversational language model with tool-augmented capabilities, offering multi-turn dialogue in Chinese and English. The project provides its training data and fine-tuning code alongside the model weights.
Project overview
By open-sourcing 300,000 plugin-augmented multi-turn dialogue samples and full supervised fine-tuning code, MOSS enables developers to build and customize local tool-using assistants rather than relying solely on closed proprietary APIs.
Project type
AI Agent · Model Development · Model Runtime
Use cases
Chat Assistants
Deployment
Refer to project documentation
License
Apache-2.0
Best for
AI engineers and researchers needing a customizable, open-source bilingual model with tool-use capabilities.
Developers looking to run a local conversational assistant with external plugin integration.
Key capabilities
Supports multi-turn dialogue, instruction following, and helpfulness in both Chinese and English.
Supports Int4 and Int8 quantized models to reduce VRAM requirements for running inference.
Includes code for supervised fine-tuning (SFT) of the MOSS base model to enable custom adaptations.
Limitations and risks
Because Triton only supports Linux and WSL, native Windows and macOS environments are unsupported.
The model may generate misleading replies with factual errors or produce harmful content containing bias and discrimination.
The Int4 and Int8 quantized models do not support model parallelism.
Distributors of the model are responsible for any negative consequences arising from the generation of harmful content.
Getting started
Clone the repository, create a conda environment with Python 3.8, and install dependencies via pip install -r requirements.txt.
Ensure access to significant GPU hardware (FP16 requires 31GB VRAM, Int8 requires 16GB VRAM, Int4 requires 7.8GB VRAM) and a Linux or WSL system, as native Windows and macOS are not supported.
Download the necessary MOSS family model weights and datasets from the Hugging Face Hub before running the application.
Launch a local session by executing a demo script, such as python moss_cli_demo.py.
Evidence and sources
GitHub project description: An open-source tool-augmented conversational language model from Fudan University