AI with LLM8850

In this video I explore the LLM8850 AI module from M5Stack. M5Stack produce mainly esp32 based components and there are two product Raspberry Pi products from them. The other is the CM4Stack that utilizes the Raspberry Pi CM4 module. Despite their lack of experience with the Raspberry Pi M5Stack have produce a few very successfully and powerful AI modules and most of these are driven by Linux.

The LLM8850 has a very high performance with a NPU (Neural Processor Unit) performing at a peak of 24 TOPS @ INT8 and in addition it has a octa-core  Cortex‑A55 1.7 GHz CPU.

I installed the LLM8850 on a Raspberry Pi IO Board in the M.2 socket and the system is driven by a Raspberry Pi CM5.
For software installation I install all of drivers for LLM8850 and the demos from the M5Stack website. I was latter to discover that a singlew system update to the operatin system corrupted my links and on the second day none of my examples and demos worked. This will be explored in a second video.

Installation went well and I ran many of the vision demos including:
Yolo11-Pose
Yolo11-seg
Yolo11-faces
All of these worked and are included in the video

I had trouble on the second day of trials after a system update and that corrupted the LLM8850 drivers. At first I could not figure out what was the problem. None of the demos and examples worked so I sought help from Microsoft Copilot. It gave me the clue. Likely that the drivers are no longer compatible with the updated system OS kernel. After uninstalling the drivers and they reinstalling them everything worked.
I needed to take care when executing some of the demo examples to make sure I was in the correct file directory.

I still have to test the following video AI examples:
MixFormerV2: improved unified Transformer-based object tracking model that enhances tracking accuracy and speed
Real-ESRGAN: deep learning-based image super-resolution model. It removes noise in the image and sharper the imae.
SuperResolution: real-time video frame interpolation model based
RIFE: It is used in video smoothness enhancement and animation production. It is used to double the frame rate of a video/

After that I will test and demo the LLMs (Large language models) including:
Qwen3-0.6B
Qwen3-1.7B
Qwen2.5-0.5B
DeepSeek-R1-Distill-Qwen-1.5B-GPTQ-Int4
MiniCPM4-0.5B
Multi-models models
InternVL3-1B
Qwen2.5-VL-3B-Instruct
Qwen3-VL-2B-Instruct
Qwen3-VL-4B-Instruct-GPTQ-Int4
SmolVLM2-500M-Video-Instruct
CLIP
Audio Models:
Whisper
MeloTTS
SenseVoice
CosyVoice2
3D-Speaker-MT
Generative Models:
lcm-lora-sd
SD1.5-LLM8850
LivePortrait
Applications:
Frigate. Our security system already uses this and announces objects that it detects within a defined array to our SONOS.
Immich
CosyVoice2-API
sherpa-onnx

There plenty here to keep me going and I will report my experiences and reflections in future videos.

References:

M5Stack Shop: https://shop.m5stack.com/products/ai-...)
M5Stack wiki: https://docs.m5stack.com/en/ai_hardwa...
M5Stack LLM-8850 User Guide: https://docs.m5stack.com/en/guide/ai_...