An open 13B LLM that tops its Chinese benchmark tables and fits on a 3090

What it does

Baichuan-13B is a 13-billion-parameter large language model developed by Baichuan Intelligent Technology. It ships as both a raw pretrained base (Baichuan-13B-Base) and a conversation-tuned chat model (Baichuan-13B-Chat). Both checkpoints are bilingual in Chinese and English and are distributed through Hugging Face and ModelScope. The project also publishes int8 and int4 quantized weights so the model can run on consumer-grade hardware like an Nvidia RTX 3090.

The interesting bit

The model was trained on 1.4 trillion tokens, which the authors claim is 40% more data than LLaMA-13B and the largest corpus for any open 13B model at the time. Rather than simply scaling up Baichuan-7B, the architecture switches from RoPE to ALiBi positional encoding and expands the hidden dimension to 5,120 across 40 layers. The licensing is also a selling point: academic use is completely unrestricted, and commercial use is free after a simple email application.

Key highlights

• Tops the provided C-Eval, MMLU, and CMMLU tables among comparable 13B open-source models, including Vicuna-13B and Chinese-LLaMA-Plus-13B.
• Ships both base and chat checkpoints, plus int8/int4 quantization, with the explicit goal of lowering deployment costs.
• Uses a 64,000-token vocabulary and a 4,096-token context window with ALiBi position encoding.
• Free commercial license available upon email request; academic use is unrestricted.

Caveats

• The README notes that Baichuan 2 (7B and 13B) has already been released, so this repository is effectively a predecessor.
• Benchmark claims are self-reported 5-shot results against a specific subset of competitors; independent verification is not shown.
• The model requires trust_remote_code=True to load via Transformers, indicating custom modeling code that departs from standard Hugging Face primitives.

Verdict

Worth a look if you need a commercially permissive, Chinese-capable 13B model that runs on modest hardware. Skip it if you are already using Baichuan 2 or prefer to avoid custom modeling code.