AquilaMoE-SFT

BAAI/AquilaMoE-SFT

Publisher北京智源人工智能研究院（BAAI）Verified

Download access requirementsMirror available

AquilaMoE：基于Scale-Up与Scale-Out策略的高效MoE模型训练 AquilaMoE是由北京人工智能研究院（BAAI）语言基础模型与软件团队开发的双语8×16B混合专家（Mixture of Experts，MoE）语言模型。该模型采用创新的EfficientScale训练方法，通过两阶段流程优化性能并最大程度减少数据需求。 AquilaMoE是一款双语（中英文）大语言模型，采用MoE架构，共8个专家模块，每个token激活2个专家，约使用30B参数。模型

Text Chat

View all download instructions Mirror repository documentation and license ↗

Model details

模型介绍

中文整理

AquilaMoE：基于Scale-Up与Scale-Out策略的高效MoE模型训练

Model use case

AquilaMoE是由北京人工智能研究院（BAAI）语言基础模型与软件团队开发的双语8×16B混合专家（Mixture of Experts，MoE）语言模型。该模型采用创新的EfficientScale训练方法，通过两阶段流程优化性能并最大程度减少数据需求。

主要能力

AquilaMoE是一款双语（中英文）大语言模型，采用MoE架构，共8个专家模块，每个token激活2个专家，约使用30B参数。模型支持4096个token的上下文长度。EfficientScale方法包含三个关键阶段：准备阶段、Scale-Up阶段和Scale-Out阶段，实现从小型密集模型到大型MoE模型的高效知识迁移。

输入输出

输入：最长4096个token的中文或英文文本

输出：文本生成

运行要求

该模型基于Aquila3 8×16B配置训练，具体参数如下：

上下文长度：4096

QKV偏置：启用

层数：40

隐藏层维度：5120

中间层维度：20480

KV组：8

训练token数：545B

学习率：1.5e-4

批量大小：24M

基本用法

AquilaMoE基础模型可通过微调获得优异性能。AquilaMoE-SFT为监督微调后的版本，在多项基准测试中表现优异。在AlpacaEval 2.0评估中，AquilaMoE-SFT得分21.1，优于GPT 3.5 Turbo（2023年11月6日版本，得分19.3）和GPT 3.5 Turbo（2023年3月1日版本，得分18.1）。

限制

AquilaMoE Instruct模型仅作为基础模型可轻松微调以获得优异性能的快速演示，未包含任何内容审核机制。团队期待与社区合作，探索如何使模型更好地遵守安全准则，以满足需要内容审核的部署环境需求。

License

AquilaMoE项目基于Apache 2.0许可证；AquilaMoE系列模型基于BAAI Aquila模型许可证协议。

优先采用原作者提供的中文模型卡片；仅有英文说明时由闲社AI翻译整理。重要参数、许可证和商用范围请以原文为准。

你将获得什么

完整仓库方式包含该固定版本中的权重、配置、分词器及说明文件；指定版本方式只下载所选文件或整组分片。多模态模型的投影文件、配置等依赖，请按作者说明配齐。

完整仓库可能包含多个权重格式，因此需要的磁盘空间可能大于单一格式的模型大小。下载文件不等于完成模型部署。

使用前须知

运行环境

运行模型需要的设备、工具与依赖，以原作者说明为准。

授权范围

是否允许商用、修改和分发，请查看模型许可证。

闲社提供资料与镜像下载方法，不代表已运行模型或完成安全审计。下载的代码应先检查，再执行。

Problem with the information or download?

Pinned version · Multiple download options

AquilaMoE-SFT model download

Download the full repository or choose a version or file.

Download tools · More than GGUF

Use the Hugging Face download tool to obtain BAAI/AquilaMoE-SFT. Download Safetensors, GGUF, ONNX, configuration and other repository files. Downloading does not mean the tool can run this model.

File manifest available for the pinned version
Files: 61
Repository formats: json / safetensors
Precision in filenames: not explicitly marked
Supporting files: documentation or licenses, configuration, and tokenizer files

This information comes from the file manifest. Not every file is a separate model; weights have not necessarily been downloaded, verified or tested.

Choose a mirror or official source, your operating system and download scope below, then copy the command. Install the download tool first if needed.

Pinned version: 564e1e9e867ffa4380498d72b3d6010d6f17e71e. Downloads files only; does not start the model.

Choose a download source

已根据访问地区推荐 Hugging Face 官方线路，你可以手动切换。

Choose what to download

Need just one GGUF precision? Load the file manifest and choose a version. Direct file links do not require Python; the command below can download multiple shards together.

Quantization / file

This reads the pinned version's file manifest; no model weights are downloaded.

Choose your operating system

For command-line downloads, follow these steps to get the full repository or selected files. Direct file links do not need these steps.

Open a terminal window

Create and open a folder for the model. In the folder's address bar, enter powershell and press Enter.

If Python is not installed, get it from the official Python website. In the terminal, enter python --version; continue when a version number appears.

First time? Install the download tool

Copy this command, paste it into the terminal, press Enter and wait for installation to finish.

python -m pip install --upgrade huggingface_hub

Already installed? Continue to step 3.

Run the download command

Select the button below, paste into the same terminal and press Enter. The tool downloads all files included in the current command.

Copying does not start a download. Paste and run the command on your device.

Where are the downloaded files?

Open the folder you chose. Look for the models folder. When progress stops and the terminal returns to its prompt without errors, check that all files are present. 'Copied' does not mean 'downloaded'.

No response or an error?

python not found: Install Python and reopen the terminal. On Windows, follow the installer's instructions to enable command-line access.

hf not found: Confirm that step 2 succeeded, then reopen the terminal. If needed, add Python's Scripts folder to your user PATH. Do not repeatedly download scripts.

Network interrupted: Stay in the same folder and run the same command again. The tool checks existing files.

Permission or disk error: Choose a writable folder with enough free space. Do not disable system security protection.

View the full download command

$env:HF_ENDPOINT = "https://huggingface.co"
$env:HF_HUB_DISABLE_IMPLICIT_TOKEN = "1"
$env:HF_HUB_DISABLE_XET = "1"
hf download "BAAI/AquilaMoE-SFT" --revision "564e1e9e867ffa4380498d72b3d6010d6f17e71e" --local-dir "./models/BAAI--AquilaMoE-SFT"

Files transfer directly from the source you choose, not through Xianshe servers.

The version is pinned. Repeating the same command checks existing files. Change the source manually if it is slow or unavailable.

Public models do not require submitting keys to Xianshe. Restricted models still require the author's authorization. On Linux / macOS, use python3 instead of python when appropriate.