Model reference · open weights

Baichuan2-4bits

LLMs baichuan-inc Text gen 1 build Its own licence terms 215 dl/mo

Baichuan2-4bits is an open-weight language model from baichuan-inc. Baichuan2-13B-Chat-4bits (BF16) weighs 9.1 GB; the smallest configuration that runs it is 2× RTX 3060 12 GB.

  • Baichuan2-4bits is a 13B parameter text-generation model developed by baichuan-inc, available as a 4-bit quantized version of the Baichuan2-13B-Chat model.
  • It supports English and Chinese and is trained on a corpus of 2.6 trillion tokens.
  • The model requires a PyTorch 2.0 environment for inference and is released under an "other" license that permits commercial use for entities with fewer than 1 million daily active users upon approval.

Summary of the baichuan-inc/Baichuan2-13B-Chat-4bits model card, 2026-10-01

What it is

Released bybaichuan-inc
Released2023-08-30
VRAM9.1 GB for the weights

What it runs on

Memory and cards for Baichuan2-13B-Chat-4bits (BF16)

9.1 GBweights, file size
819 MBcache per 1K tokens
754 MBruntime overhead, at least
CardRequests at onceContext maxMemory
8K each32K each
RTX 3060 12 GB——2K11.6 GB
RTX 4060 Ti 16 GB——6K15.4 GB
RTX 3090 24 GB2—16K23.4 GB
RTX 4090 24 GB2—16K23.4 GB
RTX 5090 32 GB3—25K31.0 GB
L40S 48 GB5140K44.0 GB
A100 80 GB10281K78.2 GB
H100 80 GB9276K78.1 GB
RTX PRO 6000 Blackwell 96 GB11295K93.8 GB
DGX Spark (GB10) 128 GB unified133111K107 GB
H200 141 GB184128K138 GB
B200 180 GB246128K176 GB
2× RTX 3060 12 GB
tensor parallel
1—15K11.6 GB a card
2× RTX 4060 Ti 16 GB
tensor parallel
3—24K15.4 GB a card
2× L40S 48 GB
tensor parallel
11292K44.0 GB a card
4× RTX 4090 24 GB
tensor parallel
12396K23.4 GB a card
4× RTX 3090 24 GB
tensor parallel
12397K23.4 GB a card
4× RTX 5090 32 GB
tensor parallel
164128K31.0 GB a card
Memory needed at each load
Requests at once8K tokens each32K tokens each
116.5 GB36.7 GB
543.4 GB144 GB
863.5 GB225 GB
16117 GB439 GB
32225 GB869 GB
64439 GB1728 GB

One card, with vLLM's small-card settings.

From the model card

What baichuan-inc says about Baichuan2-4bits

Read the model card

Baichuan 2

百川API支持搜索增强和192K长窗口,新增百川搜索增强知识库、限时免费! 🚀 百川大模型在线对话平台 已正式向公众开放 🎉

目录/Table of Contents

更新/Update

[2023.12.29] 🎉🎉🎉 我们发布了 Baichuan2-13B-Chat v2 版本。其中:

  • 大幅提升了模型的综合能力,特别是数学和逻辑推理、复杂指令跟随能力。
  • 使用时需指定revision=v2.0,详细方法参考快速开始

模型介绍/Introduction

Baichuan 2 是[百川智能]推出的新一代开源大语言模型,采用 2.6 万亿 Tokens 的高质量语料训练,在权威的中文和英文 benchmark 上均取得同尺寸最好的效果。本次发布包含有 7B、13B 的 Base 和 Chat 版本,并提供了 Chat 版本的 4bits 量化,所有版本不仅对学术研究完全开放,开发者也仅需[邮件申请]并获得官方商用许可后,即可以免费商用。具体发布版本和下载见下表:

Baichuan 2 is the new generation of large-scale open-source language models launched by Baichuan Intelligence inc.. It is trained on a high-quality corpus with 2.6 trillion tokens and has achieved the best performance in authoritative Chinese and English benchmarks of the same size. This release includes 7B and 13B versions for both Base and Chat models, along with a 4bits quantized version for the Chat model. All versions are fully open to academic research, and developers can also use them for free in commercial applications after obtaining an official commercial license through email request. The specific release versions and download links are listed in the table below:

Base ModelChat Model4bits Quantized Chat Model
7BBaichuan2-7B-BaseBaichuan2-7B-ChatBaichuan2-7B-Chat-4bits
13BBaichuan2-13B-BaseBaichuan2-13B-ChatBaichuan2-13B-Chat-4bits

快速开始/Quick Start

在Baichuan2系列模型中,我们为了加快推理速度使用了Pytorch2.0加入的新功能F.scaled_dot_product_attention,因此模型需要在Pytorch2.0环境下运行。

In the Baichuan 2 series models, we have utilized the new feature F.scaled_dot_product_attention introduced in PyTorch 2.0 to accelerate inference speed. Therefore, the model needs to be run in a PyTorch 2.0 environment.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation.utils import GenerationConfig
tokenizer = AutoTokenizer.from_pretrained("baichuan-inc/Baichuan2-13B-Chat-4bits",
    revision="v2.0",
    use_fast=False,
    trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("baichuan-inc/Baichuan2-13B-Chat-4bits",
    revision="v2.0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True)
model.generation_config = GenerationConfig.from_pretrained("baichuan-inc/Baichuan2-13B-Chat-4bits", revision="v2.0")
messages = []
messages.append({"role": "user", "content": "解释一下“温故而知新”"})
response = model.chat(tokenizer, messages)
print(response)
"温故而知新"是一句中国古代的成语,出自《论语·为政》篇。这句话的意思是:通过回顾过去,我们可以发现新的知识和理解。换句话说,学习历史和经验可以让我们更好地理解现在和未来。

这句话鼓励我们在学习和生活中不断地回顾和反思过去的经验,从而获得新的启示和成长。通过重温旧的知识和经历,我们可以发现新的观点和理解,从而更好地应对不断变化的世界和挑战。

注意:如需使用老版本,需手动指定revision参数,设置revision=v1.0

Benchmark 结果/Benchmark Evaluation

我们在[通用]、[法律]、[医疗]、[数学]、[代码]和[多语言翻译]六个领域的中英文权威数据集上对模型进行了广泛测试,更多详细测评结果可查看[GitHub]。

We have extensively tested the model on authoritative Chinese-English datasets across six domains: General, Legal, Medical, Mathematics, Code, and Multilingual Translation. For more detailed evaluation results, please refer to GitHub.

7B Model Results

C-EvalMMLUCMMLUGaokaoAGIEvalBBH
5-shot5-shot5-shot5-shot5-shot3-shot
GPT-468.4083.9370.3366.1563.2775.12
GPT-3.5 Turbo51.1068.5454.0647.0746.1361.59
LLaMA-7B27.1035.1026.7527.8128.1732.38
LLaMA2-7B28.9045.7331.3825.9726.5339.16
MPT-7B27.1527.9326.0026.5424.8335.20
Falcon-7B24.2326.0325.6624.2424.1028.77
ChatGLM2-6B50.2045.9049.0049.4445.2831.65
[Baichuan-7B]42.8042.3044.0236.3434.4432.48
[Baichuan2-7B-Base]54.0054.1657.0747.4742.7341.56

13B Model Results

C-EvalMMLUCMMLUGaokaoAGIEvalBBH
5-shot5-shot5-shot5-shot5-shot3-shot
*

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms