MOSS

MOSS is a conversational language model developed by the natural language processing lab at Fudan university, named after the self-aware AI in the film The Wandering Earth. It drew wIDE attention as one of the earliest Chinese-built GPT-style chat systems offered for public Testing, and the team later published open weights so researchers and developers could build on the model instead of recreating the underlying Technology from scratch.

What it does

MOSS is a dialogue-oriented model trained on a large corpus of Chinese, English and code text. The released base model cARRies roughly 16 billion parameters, and the team has described it as light enough to run on a single consumer graphics card when precision requirements are relaxed. It handles multi-turn conversation, code writing and math-style reasoning, and it can call on external tools such as a search engine, an image generator and a calculator through a plugin mechanism.

Key capabilities

• Bilingual chat: supports conversation in both Chinese and English across multiple turns.
• tool use: plugin support for search, text-to-Image Generation, calculators and equation solving.
• Open releases: base pretrained weights plus instruction-tuned variants published on model hosting platforms.
• Partial data release: a subset of training data was made available, with the full set described as coming later.
• Source tracing: dEMOnstration examples show the model citing where its Information came from.

Who it is for

MOSS suits academic researchers studying dialogue models and instruction tuning, developers who want a bilingual foundation to fine-tune for domain-specific assistants, and teams that prefer open weights over a closed API. It is also useful for students and hobbyists experimenting with local inference on modest hardware, though results depend heavily on quantization and setup.

How to get started

Start by reviewing the project repositories, where model cards describe the available checkpoints and their intended use. Download the weights through the linked model hosting pages, then load them with a compatible Inference Framework. Expect to configure tokenizer files, context length and generation settings yourself; this is a developer-oriented release rather than a polished consumer app. A web demo has been offered for quick trials, so check whether it is currently online before relying on it. Because the model is large, plan for memory and compute requirements, and test with short Prompts first.

Pricing and access caveat

The open-source releases are free to download and use under the terms stated in the project license, but commercial use conditions, availability of the full final model, and the completeness of the training data release may vary and should be verified on the official pages. The public demo may queue, change, or go offline without notice. PerfORMance in Chinese was acknowledged by the team as liMITed in earlier iterations, so evaluate ouTPUt quality for your own use case rather than assuming parity with larger commercial systems.