MaixPy 4.12.5

Sipeed Maix Vision Python SDK

MaixPy (v4)

Let's Sipeed up, Maximize AI's power!

MaixPy (v4): Easily create AI projects with Python on edge device

Quick Start | Documentation | API | Hardware

English | 中文

Feature Overview

MaixPy offers simple Python programming combined with powerful edge computing hardware. Integrated hardware peripheral operations, video streaming, AI vision algorithms, audio algorithms, and LLM / VLM etc. With its plug-and-play design, MaixPy enables you to quickly implement your intelligent projects.

Additionally, MaixPy provides the MaixVision IDE, MaixHub online training platform, detailed documentation, and even a C/C++ SDK with identical APIs, ensuring seamless development and production deployment.

Here is a partial video demonstration of the features. For more documentation, please visit the official website: wiki.sipeed.com/maixpy/

If you like this project, please click Star on the top right of the MaixPy Project to encourage us to develop more exciting content!

Concise and Efficient Code (API) Design

With MaixPy you can easily create AI vision project within 10 lines of code:

from maix import camera, display, image, nn

classifier = nn.Classifier(model="/root/models/mobilenetv2.mud")
cam = camera.Camera(classifier.input_width(), classifier.input_height(), classifier.input_format())
disp = display.Display()

while 1:
    img = cam.read()
    res = classifier.classify(img)
    max_idx, max_prob = res[0]
    msg = f"{max_prob:5.2f}: {classifier.labels[max_idx]}"
    img.draw_string(10, 10, msg, image.COLOR_RED)
    disp.show(img)

Result:

Edge(embeded) friendly

Simply use hardware peripheral like serial port:

from maix import uart

devices = uart.list_devices()

serial = uart.UART(devices[0], 115200)
serial.write_str("hello world")
print("received:", serial.read(timeout = 2000))

MaixVision workstation

We also provide a handy MaixVision workstation software to make development easier and faster:

MaixVision

MaixHub online platform

MaixHub provide free online AI train service, one click to train AI model and deploy to MaixCAM even you have no AI knowledge and expensive training equipment.

Hardware platform MaixCAM

And we provide two powerful hardware platform MaixCAM2, MaixCAM and MaixCAM-Pro.

Performance comparison

K210 and v831 are outdated, they have many limitations in memory, performance, NPU operators missing etc.
No matter you are using them or new comer, it's recommended to upgrade to MaixCAM and MaixPy v4.

Here's the comparison between them:

Feature	Maix-I K210	MaixCAM	MaixCAM2
CPU	400MHz RISC-V x2	1GHz RISC-V(Linux) 700MHz RISC-V(RTOS) 25~300MHz 8051(Low Power)	1.2GHz A53 x2(Linux) RISC-V 32bit E907(RTT)
Memory	6MB SRAM	256MB DDR3	1GB / 4GB LPDDR4
NPU	0.25Tops@INT8 official says 1T but...	1Tops@INT8	3.2Tops@INT8
Encoder	❌	2880x1620@30fps H.254/H.265/JPEG	3840*2160@30fps H.254/H.265/JPEG
Decoder	❌	2880x1620@30fps H.264/JPEG	1080p@60fps H.264/JPEG
Screen	2.4" 320x240	2.3" 552x368(MaixCAM) 2.4" 640x480(MaixCAM-Pro) 5" 1280x720 7" 1280x800 10“ 1280x800	2.4" 640x480 5" 1280x720 7" 1280x800 10“ 1280x800
Touchscree	❌	2.3" 552x368/2.4" 640x480	2.4" 640x480
Camera	30W	500W(5M)	800W(8M)
AI ISP	❌	❌	✅
WiFi	2.4G	WiFi6 2.4G/5G	WiFi6 2.4G/5G
BLE	❌	BLE5.4	BLE5.4
USB	❌	USB2.0	USB2.0
Ethernet	❌	100M(Optional)	100M(on board FPC, can convert to RJ45 module)
SD Card	SPI	SDIO	SDIO
OS	RTOS	Linux(BuildRoot) + RTOS	Linux(Ubuntu) + RTT
Porgraming Language	C / C++ / MicroPython	C / C++ / Python3	C / C++ / Python3
SDK	MaixPy-v1	MaixCDK + MaixPy v4 + opencv + numpy + ... Pure Python package or cross-compile manually	MaixCDK + MaixPy v4 + opencv + numpy + scipy + ... Many AArch64 pre-compiled packages, and support compile on board
PC Software	MaixPy IDE	MaixVision Workstation	MaixVision Workstation
Documentation	⭐️⭐️⭐️⭐️	🌟🌟🌟🌟🌟	🌟🌟🌟🌟🌟
Online AI train	⭐️⭐️⭐️	🌟🌟🌟🌟🌟	🌟🌟🌟🌟🌟
Official APPs	⭐️	🌟🌟🌟🌟🌟	🌟🌟🌟🌟🌟
Ease of use	⭐️⭐️⭐️⭐️	🌟🌟🌟🌟🌟	🌟🌟🌟🌟🌟
AI classify(224x224)	MobileNetv1 50fps MobileNetv2 ❌ Resnet ❌	MobileNetv2 130fps Resnet18 62fps Resnet50 28fps	MobileNetv2 1218fps Resnet50 200fps
AI detect only forward part / [include pre-post process parts(Python)] / [dual buff mode(Python)]	YOLOv2: 224x224: 15fps	YOLOv5s: 224x224: 100fps 320x256 70fps 640x640: 15fps YOLOv8n: 640x640: 23fps YOLO11n: 224x224: 175fps 320x224: 120fps 320x320: 95fps 640x640: 23fps	YOLOv5s: 224x224: 495fps 320x256: 400fps 640x480: 106fps / 73fps / 103fps 640x640: 80fps YOLO11n: 224x224: 1214fps 640x480: 168fps / 77fps / 143fps 640x640: 113fps / 56fps / 98fps YOLO11s: 640x480: 87fps / 53fps / 83fps 640x640: 62fps / 39fps / 59fps YOLO11l: 640x640: 19fps / 16fps / 19fps
LLM	❌	❌	Qwen/DeepSeek 0.5B(fftf: 640ms, 9 tokens/s) Qwen/DeepSeek 1.5B(fftf: 1610ms, 4 tokens/s) VLM(InterVL 1B) Mode models
OpenMV algorithms	test image refer to Benchmark APP	test image refer to Benchmark APP test date: 2025.8.22，update may have optimization	test image refer to Benchmark APP test date: 2025.8.22，update may have optimization
Binary	Gray 320x240: 7.4ms (135fps) Gray 640x480: ❌ RGB 320x240: 11.3ms (88.5fps) RGB 640x480: ❌	Gray 320x240: 3.1ms (326fps) Gray 640x480: 11ms (90fps) RGB 320x240: 13.2ms (75fps) RGB 640x480: 52.8ms (18fps)	Gray 320x240: 1.3ms (799fps) Gray 640x480: 4.8ms (206fps) RGB 320x240: 3.4ms (294fps) RGB 640x480: 13.3ms (75fps)
Find blobs	320x240: 8.8ms (114fps) 640x480: ❌	320x240: 7ms (143fps) 640x480: 20ms (50fps)	320x240: 3.7ms (271fps) 640x480: 11.1ms (89fps)
1channel histogram	320x240: 7.7ms (130fps) 640x480: ❌	320x240: 10.9ms (91fps) 640x480: 42.8ms (23fps)	320x240: 1.5ms (661fps) 640x480: 5.9ms (168fps)
QR Code	320x240: 130.8ms (7.6fps) 640x480: ❌	640x480: 136.9ms (7fps) NPU acceleration：   320x240: 22.1ms (45fps)   640x480: 57.6ms (17fps)	640x480: 57.9ms (17fps) NPU acceleration：   320x240: 9.2ms (109fps)   640x480: 23.2ms (43fps)
OpenCV algorithms		test image refer to Benchmark APP test date: 2025.8.22，update may have optimization	test image refer to Benchmark APP test date: 2025.8.22，update may have optimization
Binary	❌	Gray 320x240: 2.2ms (463fps) Gray 640x480: 7.1ms (140fps)	Gray 320x240: 0.1ms (8174fps) Gray 640x480: 0.3ms (2959fps)
Gray adaptive binary	❌	320x240: 5.8ms (171fps) 640x480: 21.3ms (46fps)	320x240: 1.6ms (608fps) 640x480: 6.3ms (159fps)
1channel histogram	❌	320x240: 1ms (1000fps) 640x480: 6.2ms (160fps)	320x240: 0.4ms (2308fps) 640x480: 1.7ms (604fps)
Find Contours	❌	320x240: 2.8ms (351fps) 640x480: 8.6ms (116fps)	320x240: 0.4ms (2286fps) 640x480: 1.4ms (692fps)

Maix Ecosystem

MaixPy not only a Python SDK, but have a whole ecosystem, including hardware, software, tools, docs, even cloud platform etc. See the picture below:

Who are using MaixPy?

• AI Algorithm Engineer who want to deploy your AI model to embedded devices.

MaixPy provide easy-to-use API to access NPU, and docs to help you develop your AI model.

• STEM teacher who wants to teach AI and embedded devices to students.

MaixPy provide easy-to-use API, PC tools, online AI train service ... Let you focus on teaching AI, not the hardware and complicated software usage.

• Maker who want to make some cool projects but don't want to learn complicated hardware and software.

MaixPy provide Python API, so all you need is learn basic Python syntax, and MaixPy's API is so easy to use, you can make your project even in a few minutes.

• Engineer who want to make some projects but want a prototype as soon as possible.

MaixPy is easy to build projects, and provide corresponding C++ SDK, so you can directly use MaixPy to deploy or transfer Python code to C++ in a few minutes.

• Students who want to learn AI, embedded development.

We provide many docs and tutorials, and lot of open source code, to help you find learning route, and grow up step by step. From simple Python programming to Vision, AI, Audio, Linux, RTOS etc.

• Enterprise who want to develop AI vision products but have no time or engineers to develop complicated embedded system.

Use MaixPy even graphic programming to develop your products with no more employees and time. For example, add a AI QA system to your production line, or add a AI security monitor to your office as your demand.

• Contestants who want to win the competition.

MaixPy integrate many functions and easy to use, fasten your work to win the competition in limited time. There are already many contestants win the competition with MaixPy.

Compile Source Code

If you want to compile MaixPy firmware from source code, refer to Build MaixPy source code page.

License

All files in this repository are under the terms of the Apache License 2.0 Sipeed Ltd. except the third-party libraries or have their own license.

Community

• Project sharing: maixhub.com/share
• App sharing: maixhub.com/app
• Discussion: maixhub.com/discussion
• QQ group: 862340358
• Telegram: t.me/maixpy
• Github issues: github.com/sipeed/maixpy/issues