#2222nd of 34 in Open models
moss
Bilingual Chinese-English 16B chat model with plugins, data and fine-tuning code
- Stars
- 12.3k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Language
- Python
11,987 on 31 Mar 202612,283 on 20 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
MOSS-003 is a 16B-parameter Chinese/English conversational language model from Fudan University, released as base, SFT and plugin-tuned checkpoints on Hugging Face. The repo includes the roughly 1.1M-dialogue SFT data, Transformers inference examples, GPTQ int4/int8 variants and fine-tuning code. The plugin variant calls web search, calculator, equation solver and text-to-image through a text command format.
Who it is for: Researchers and engineers running or tuning a Chinese/English 16B chat model
Strengths
- Int4 checkpoints run on one GPU at about 12GB VRAM for a single turn
- Releases SFT conversation data, including about 300k plugin-augmented dialogues
- Plugin model emits explicit Search/Calculate/Solve/Text2Image commands you wire up yourself
- Includes both single-GPU and multi-GPU (accelerate) inference examples
Weaknesses
- Model weights are AGPL-3.0 and data is CC BY-NC 4.0, unlike the Apache-2.0 code
- FP16 needs about 31GB to load and 81GB at the 2048-token maximum
- Quantized models run on a single card only and need triton, so Linux or WSL only
- Preference-tuned moss-moon-003 models are listed as coming soon, not released
What it needs
- GPU required
- Needs PyTorch, transformers, accelerate, triton (quantized models), conda
- Models: moss-moon-003-base, moss-moon-003-sft, moss-moon-003-sft-plugin, GPTQ int4, GPTQ int8
Also in Open models
See all 34| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 87 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 84 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 75 out of 100 |
| #4 | speechPyTorch toolkit for training and deploying ASR, TTS and speech LLM models | 70 out of 100 |
| #5 | yueLyrics-to-song model that writes an editable melody and chord score first | 65 out of 100 |