lonePatient/awesome-pretrained-chinese-nlp-models

What this project is

A curated, continuously updated list (an "Awesome" list) of open-source Chinese NLP models — including large language models, multimodal models, and classic pretrained models — plus the tools, benchmarks, and platforms around them.

What it contains

1. Large language models (LLMs), organized into six categories:

Category	What it covers
Base-LLM	General-purpose foundation models with more than 7 billion parameters (e.g., Qwen-2.5, InternLM, Llama-3.1)
Domain-Base-LLM	Foundation models for specific fields — finance, medical, code, math (e.g., Tongyi-Finance, ChiMed-GPT, CodeShell)
ChatLLM	Conversational models for dialogue and Q&A (e.g., DeepSeek-V3, GLM-4.6, Kimi-K2)
Domain-ChatLLM	Conversational models for vertical domains, mostly code-focused (e.g., DeepSeek-Coder-V2, CodeGeeX4)
MultiModal-ChatLLM	Models handling images, audio, video, 3D, or documents alongside text (e.g., MiniMax-Music3, Qianfan-OCR)
ReasoningLLM	Models with strong math/logic reasoning ability (e.g., GLM-5.3, Qwen3.8-Flash-Next)

2. Classic pretrained model series (pre-LLM era):

• NLU (natural language understanding): BERT, RoBERTa, ALBERT, ERNIE, etc. — 29 models
• NLG (natural language generation): GPT, T5, BART, CPM, RWKV, etc. — 18 models
• NLU-NLG (both): UniLM, GLM, CPT, SimBERT — 9 models
• Multimodal pretrained: WenLan, CogView, Chinese-CLIP, OFA — 13 models

3. Embedding models — 16 models for converting text/images/video into vector representations (e.g., Qwen3-Embedding, Conan-embedding-v1).

4. Evaluation benchmarks — seven tools for measuring model quality:

• C-Eval: 13,948 multiple-choice questions across 52 subjects
• FlagEval: multi-dimensional evaluation (accuracy, efficiency, robustness)
• SuperCLUElyb: anonymous human-vs-human model battles using an Elo rating system
• XiezhiBenchmark: 220,000 questions across 516 disciplines
• Open LLM Leaderboard: HuggingFace's English-focused leaderboard
• Chinese LLM Safety Evaluation Platform: checks for hate speech, bias, privacy violations, etc.
• OpenCompass: open-source evaluation framework supporting zero-shot, few-shot, and chain-of-thought testing

5. Open-source model platforms — where to download models: HuggingFace, ModelScope, flagopen, and 始智AI (wisemodel).

6. Dataset repositories — seven sources for training data, including CLUECorpus2020 (100GB of Chinese text) and MNBVC (a massive Chinese corpus).

7. Other Awesome lists — 52 additional curated resource lists (AI agents, Claude skills, etc.).

How each model entry is presented

Every model appears in a table with: size (parameter count), release date, language(s), architecture (e.g., MoE = mixture-of-experts, CD = causal decoder), download link, project link, institution, and sometimes a paper or blog link. Full lists live in separate docs/ files; the README shows only a sample.

Practical notes

• The README recommends a HuggingFace mirror (hf-mirror.com) for users in China who have trouble downloading from the main site.
• The project links to an academic survey paper, "A Survey of Large Language Models" (arXiv:2303.18223).
• The list is actively maintained — the update log shows new models added regularly (e.g., August 2026 additions like Hy4-preview and GLM-5.3).

In one sentence

A one-stop, well-organized directory of open-source Chinese NLP models — from classic BERT-era models to modern LLMs — along with the benchmarks, platforms, and datasets you need to use or evaluate them.