OpenMOSS/MOSS

What it solves

MOSS is an open-source conversational language model designed to provide a helpful, honest, and harmless AI assistant capable of fluent bilingual (Chinese and English) communication and tool use. It addresses the need for high-quality, open-weight models that can follow complex instructions and interact with external tools to provide more accurate and factual responses.

How it works

Built on a 16-billion parameter architecture, the moss-moon series is pre-trained on approximately 700 billion Chinese, English, and code tokens. The model undergoes a multi-stage training process: self-supervised pre-training, supervised fine-tuning (SFT) on over 1.1 million rounds of dialogue data, and preference training to improve stability, safety, and factuality. For tool use, specific plugin-enhanced versions are trained on 300,000 rounds of plugin-augmented dialogue data, allowing the model to generate internal thoughts and specific commands to call external APIs (such as search engines, calculators, and equation solvers) before generating a final response.

Who it’s for

It is intended for developers and researchers who want to deploy a bilingual LLM locally or fine-tune it for specific tasks. It is suitable for users with hardware ranging from a single A100/A800 GPU to consumer-grade NVIDIA 3090s (especially when using INT4/INT8 quantized versions).

Highlights

• Bilingual Proficiency: Fluent communication in both Chinese and English.
• Tool Integration: Native support for plugins including web search, calculators, equation solvers, and text-to-image generation.
• ** uma Quantization**: Available in FP16, INT8, and INT4 precisions to lower hardware barriers for local deployment.
• Open Ecosystem: Provides not only the models but also the SFT and preference datasets used for training.