Native-LLM-for-Android

Overview

Demonstration of running a native Large Language Model (LLM) on Android devices. Currently supported models include:

• Qwen3: 0.6B, 1.7B, 4B...
• Qwen3-VL: 2B, 4B...
• Qwen2.5-Instruct: 0.5B, 1.5B, 3B...
• Qwen2.5-VL: 3B
• DeepSeek-R1-Distill-Qwen: 1.5B
• MiniCPM-DPO/SFT: 1B, 2.7B
• Gemma-3-it: 1B, 4B...
• Phi-4-mini-Instruct: 3.8B
• Llama-3.2-Instruct: 1B
• InternVL-Mono: 2B
• InternLM-3: 8B
• Seed-X: PRO-7B, Instruct-7B
• HunYuan: MT-1.5-1.8B/7B

Update

• 2026/01/04：Update HunYuan-MT-1.5
• 2025/11/11：Update Qwen3-VL.
• 2025/09/07：Update HunYuan-MT.
• 2025/08/02：Update Seed-X.
• 2025/04/29：Update Qwen3.
• 2025/04/05：Update Qwen2.5, InternVL-Mono q4f32 + dynamic_axes.
• 2025/02/22：Support loading with low memory mode: Qwen, QwenVL, MiniCPM_2B_single; Set low_memory_mode = true in MainActivity.java.
• 2025/02/07：DeepSeek-R1-Distill-Qwen: 1.5B (Please using Qwen v2.5 Qwen_Export.py)

Getting Started

1. Download Models:

   • Quick Try: Qwen3-1.7B-Android

2. Setup Instructions:

   • Place the downloaded model files into the assets folder.
   • Decompress the *.so files stored in the libs/arm64-v8a folder.

3. Model Notes:

   • Demo models are converted from HuggingFace or ModelScope and optimized for extreme execution speed.
   • Inputs and outputs may differ slightly from the original models.
   • For Qwen2VL / Qwen2.5VL, adjust the key variables to match the model parameters.
     • GLRender.java: Line 37, 38, 39
     • project.h: Line 14, 15, 16, 35, 36, 41, 59, 60

4. ONNX Export Considerations:

   • It is recommended to use dynamic axes and q4f32 quantization.

Tokenizer Files

• The tokenizer.cpp and tokenizer.hpp files are sourced from the mnn-llm repository.

Exporting Models

1. Navigate to the Export_ONNX folder.
2. Follow the comments in the Python scripts to set the folder paths.
3. Execute the ***_Export.py script to export the model.
4. Quantize or optimize the ONNX model manually.

Quantization Notes

• Use onnxruntime.tools.convert_onnx_models_to_ort to convert models to *.ort format. Note that this process automatically adds Cast operators that change FP16 multiplication to FP32.
• The quantization methods are detailed in the Do_Quantize folder.

Additional Resources

• Explore more projects: DakeQQ Projects

Performance Metrics

Qwen

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	Qwen-2-1.5B-Instruct q8f32	20 token/s
Android 15	Vivo x200 Pro	MediaTek_9400-CPU	Qwen-3-1.7B-Instruct q4f32 dynamic	37 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Qwen-3-1.7B-Instruct q4f32 dynamic	18.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Qwen-2.5-1.5B-Instruct q4f32 dynamic	20.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Qwen-2-1.5B-Instruct q8f32	13 token/s
Harmony 3	荣耀 20S	Kirin_810-CPU	Qwen-2-1.5B-Instruct q8f32	7 token/s

QwenVL

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	QwenVL-2-2B q8f32	15 token/s
Harmony 4	P40	Kirin_990_5G-CPU	QwenVL-2-2B q8f32	9 token/s
Harmony 4	P40	Kirin_990_5G-CPU	QwenVL-2.5-3B q4f32 dynamic	9 token/s

惑

DeepSeek-R1

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	Distill-Qwen-1.5B q4f32 dynamic	34.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Distill-Qwen-1.5B q4f32 dynamic	20.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Distill-Qwen-1.5B q8f32	13 token/s
HyperOS 2	Xiaomi-14T-Pro	MediaTek_9300+-CPU	Distill-Qwen-1.5B q8f32	22 token/s

MiniCPM

OS	Device	Backend	Model	Inference (1024 Context)
Android 15	Nubia Z50	8_Gen2-CPU	MiniCPM4-0.5B q4f32	78 token/s
Android 13	Nubia Z50	8_Gen2-CPU	MiniCPM-2.7B q8f32	9.5 token/s
Android 13	Nubia Z50	8_Gen2-CPU	MiniCPM-1.3B q8f32	16.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	MiniCPM-2.7B q8f32	6 token/s
Harmony 4	P40	Kirin_990_5G-CPU	MiniCPM-1.3B q8f32	11 token/s

Gemma

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	Gemma-1.1-it-2B q8f32	16 token/s

Phi

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	Phi-2-2B-Orange-V2 q8f32	9.5 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Phi-2-2B-Orange-V2 q8f32	5.8 token/s

Llama

OS	Device	Backend	Model	Inference (1024 Context)
Android 13	Nubia Z50	8_Gen2-CPU	Llama-3.2-1B-Instruct q8f32	25 token/s
Harmony 4	P40	Kirin_990_5G-CPU	Llama-3.2-1B-Instruct q8f32	16 token/s

InternVL

OS	Device	Backend	Model	Inference (1024 Context)
Harmony 4	P40	Kirin_990_5G-CPU	Mono-2B-S1-3 q4f32 dynamic	10.5 token/s

MiniCPM

OS	Device	Backend	Model	Inference (1024 Context)
Android 15	Nubia Z50	8_Gen2-CPU	MiniCPM4-0.5B q4f32	78 token/s

Demo Results

Qwen2VL-2B / 1024 Context

Qwen2-1.5B / 1024 Context

概述

展示在 Android 设备上运行原生大型语言模型 (LLM) 的示范。目前支持的模型包括：

• Qwen3: 0.6B, 1.7B, 4B...
• Qwen3-VL: 2B, 4B...
• Qwen2.5-Instruct: 0.5B, 1.5B, 3B...
• Qwen2.5-VL: 3B
• DeepSeek-R1-Distill-Qwen: 1.5B
• MiniCPM-DPO/SFT: 1B, 2.7B
• Gemma-3-it: 1B, 4B...
• Phi-4-mini-Instruct: 3.8B
• Llama-3.2-Instruct: 1B
• InternVL-Mono: 2B
• InternLM-3: 8B
• Seed-X: PRO-7B, Instruct-7B
• HunYuan: MT-1.5-1.8B/7B

最近更新

• 2026/01/04：Update HunYuan-MT-1.5
• 2025/11/11：更新 Qwen3-VL。
• 2025/09/07：更新 HunYuan-MT。
• 2025/08/02：更新 Seed-X。
• 2025/04/29：更新 Qwen3。
• 2025/04/05: 更新 Qwen2.5, InternVL-Mono q4f32 + dynamic_axes。
• 2025/02/22：支持低内存模式加载: Qwen, QwenVL, MiniCPM_2B_single; Set low_memory_mode = true in MainActivity.java.
• 2025/02/07：DeepSeek-R1-Distill-Qwen: 1.5B （请使用 Qwen v2.5 Qwen_Export.py）。

入门指南

1. 下载模型：

   • Quick Try: Qwen3-1.7B-Android

2. 设置说明：

   • 将下载的模型文件放入 assets 文件夹。
   • 解压存储在 libs/arm64-v8a 文件夹中的 *.so 文件。

3. 模型说明：

   • 演示模型是从 HuggingFace 或 ModelScope 转换而来，并针对极限执行速度进行了优化。
   • 输入和输出可能与原始模型略有不同。
   • 对于Qwen2VL / Qwen2.5VL，请调整关键变量以匹配模型参数。
     • GLRender.java: Line 37, 38, 39
     • project.h: Line 14, 15, 16, 35, 36, 41, 59, 60

4. ONNX 导出注意事项：

   • 推荐使用动态轴以及q4f32量化。

分词器文件

• tokenizer.cpp 和 tokenizer.hpp 文件来源于 mnn-llm 仓库。

导出模型

1. 进入 Export_ONNX 文件夹。
2. 按照 Python 脚本中的注释设置文件夹路径。
3. 执行 ***_Export.py 脚本以导出模型。
4. 手动量化或优化 ONNX 模型。

量化说明

• 使用 onnxruntime.tools.convert_onnx_models_to_ort 将模型转换为 *.ort 格式。注意该过程会自动添加 Cast 操作符，将 FP16 乘法改为 FP32。
• 量化方法详见 Do_Quantize 文件夹。

额外资源

• 探索更多项目：DakeQQ Projects