HAOS support for M5Stack LLM-8850 M.2 Axera AX8850 AI accelerator?

[Hedda]

M5Stack LLM-8850 is a new M.2 M-Key (2242 NGFF socket) AI accelerator card based on Axera AX8850 24 TOPS @ INT8 NPU SoC:

• https://www.cnx-software.com/2025/10/03/m5stack-llm-8850-card-an-m-2-m-key-ai-accelerator-module-based-on-axera-ax8850-24-tops-soc/

Wondering if we could get support for this by default in the official Home Assistant Operating System (Linux distobution)?

• https://docs.m5stack.com/en/guide/ai_accelerator/llm-8850/m5_llm_8850_software_install

This is a new LLM8850 M.2 AI accelerator module is less expensive alternative than competes with the Hailo-8 based the Raspberry Pi AI Kit and Waveshare Hailo-8 M.2 AI Accelerator Module, and just like those this too is designed primarly for "Edge AI" on single-board-computers such as Raspberry Pi 5 / CM5, Rockchip RK3588 SBCs (like example NanoPC-T6, NanoPi R6C, Orange Pi 5 Plus, Banana Pi BPI-M7), and small-form-factor x86-64 Mini-PCs (like Intel NUC series) with a spare M.2 Key-M socket.

• https://docs.m5stack.com/en/ai_hardware/LLM-8850_Card

On paper the Axera 8850 SoC Edge NPU coprocessor is howerver far superior on general AI acceleration / AI generation and thus more flexible then the Hailo-8 M.2 cards which in turn are more or less optimized for computer vision applications.

• https://docs.m5stack.com/en/guide/ai_accelerator/overview

The M5Stack LLM-8850 card supports PCIe 2.0 ×2 lanes and onboard it has 8GB RAM (64 bit LPDDR4x @ 4266 Mbps) which NPU capable of up to 24 TOPS @ INT8 (based on Axera AX8850 Octa-Core Cortex‑A55 1.7 GHz CPU SoC). It also supports H.265/H.264 8Kp30 video encoding and 8Kp60 video decoding, with up to 16 channels for 1080p videos, so another use case would be on-the-fly video transcoding acceleration for Frigate NVR, other than real-time AI object detection (similar to the Google Coral M.2 and mini-PCIE card Edge TPU coprocessor). The main downside (other than it only having 8GB of VRAM) is that M5Stack LLM-8850 card runs a hotter so need a fan for active cooling as maximum load it uses 7W @ 3.3V and runs at 70 °C degrees (so it probably get way too hot to run inside the Home Assistant Yellow enclosure).

https://www.youtube.com/watch?v=zBlaAHn4aF8

M5Stack sells the Axera AX8850 M.2 module for $99 US-dollar in its own online store (and on its on AliExpress store):

• https://shop.m5stack.com/products/ai-8850-llm-accleration-m-2-module-ax8850

PS: M5Stack is a subsidiary of Espressif (of ESP32/ESP8266 fame):

• https://www.espressif.com/en/news/Espressif_Acquires_M5Stack

[Hedda]

FYI, Axera representative has started a Frigate NVR discussion here:

• How to contribute to the code of other edge computing AI accelerators? blakeblackshear/frigate#20283

Fork of Frigate with initial Axera support:

• https://github.com/AXERA-TECH/frigate/tree/v0.16.0-rpi-axcl

AXERA-TECH rep also mentioned these Huggingface models:

• https://huggingface.co/AXERA-TECH

  • https://github.com/AXERA-TECH

[Hedda]

FYI, also see the "Radxa AICore AX-M1" M.2 2280 M Key form-factor AI Acceleration Module (M72T) is based the same Axera AX8850 SoC:

• https://radxa.com/products/aicore/ax-m1/

  • https://radxa.com/products/aicore/

RapidAnalysis posted a demo-video of a preview unit testing DeepSeek-R1-Qwen-7B and SmolLM2-360M-Instruct large language model(s):

• https://www.youtube.com/watch?v=4dGTC-YSq1g

Radxa has a support listing for Large Language Models, small and large Vision models, Speech Models, Text-to-image Generation models:

Large Language Models:

• DeepSeek-R1-Distill
• Qwen2.5
• Qwen3
• MiniCPM4
• SmolLM3
• Llama3.2
• Gemma2
• Phi3

Vision Large Models:

• InternVL3
• Qwen2.5-VL
• SmolVLM2
• CLIP
• YOLOWorldv2

Speech Models:

• Whisper
• SenseVoice
• MeloTTS

Text-to-image Generation Model:

• Stable Diffusion v1.5

Vision Model:

• YOLO Family
• Depth-Anything-V2
• Real-ESRGAN
• MixFormerV2
• LivePortrait