How to contribute to the code of other edge computing AI accelerators?

[BUG1989]

Describe what you are trying to accomplish and why in non technical terms

• We are a company specializing in edge AI chips. Our AI chips can efficiently run excellent algorithm models such as YOLO, CLIP, Whisper, LLM, and VLM, and consume only a few watts of power.

• We have already preliminarily completed the adaptation of the M.2 computing power card LLM8850 (similar to Hailo8L, M.2242, up to 24TOPs@Int8) based on AX650N/AX8850 for the edge AI chips to Frigate, reducing the CPU usage of the target detector.

• Our next step is to further support the adaptation of JinaCLIP. Our chips have already been adapted to various CLIP models, such as CN_CLIP-L14-336, CLIP-L14-336, MobileCLIP2-S4, and have already been mass-produced on other commercial intelligent NVRs. We also hope to extend this function to the Frigate platform, so that users of Frigate can have more options.

• We can provide adaptation services for x86 platforms and aarch64 platforms (such as Raspberry Pi 5), and once completed, the source code will be made open-source.

Describe the solution you'd like

• If we want to add code to the Frigate mainline to support our chips, do you have any rules to follow? We hope to receive your guidance.

• We hope to receive guidance from the official engineers of Frigate, so that our co-processors can exert their greater capabilities and enhance the power of Frigate.

Describe alternatives you've considered
None

Additional context

• Specification Introduction of LLM8850
• We forked a version of v0.16 and made some modifications to it.
• Huggingface Model Zoo

[hawkeye217]

Support for object detection hardware can be added to Frigate through the community supported boards project. See the official documentation here: https://docs.frigate.video/development/contributing-boards

You can see examples in the codebase of other community supported boards. Here is a PR for a recently added one: #19680

You can follow the conventions in the PR and in the official documentation to submit a pull request.

[inkeep-ai-assistant[bot]]

To contribute support for new edge AI accelerators or detectors to Frigate, you should follow the standard process for contributing to the Frigate main code base:

1. Fork the main Frigate repository to your own GitHub profile and clone it locally(1).
2. Set up your development environment using the provided Docker-based workflow to ensure consistency. This includes using GNU make and Docker (with buildx), and optionally having a supported AI detector device for development(1).
3. Make your changes to add support for your edge AI accelerator. Typically, this would involve creating a new detector type or integrating your hardware into the detectors section, following the structure used for detectors like edgetpu, hailo8l, onnx, openvino, rknn, and tensorrt(2). Ensure your integration aligns with how existing detectors are structured and configured.
4. Place a test configuration at config/config.yml for local development and validation(1).
5. Test thoroughly, including unit tests (npm run test for web components), and ensure your changes do not break existing functionality. You should strive to increase test coverage(1).
6. Open a pull request to the main Frigate repository, describing your changes and providing relevant documentation and usage instructions(1).

For large or architectural changes, it is recommended to open a discussion on GitHub before submitting significant modifications(1).

You can find more details and guidelines in the official contributing guide here: Contributing To The Main Code Base(1).

📚 Sources:

1. Contributing To The Main Code Base
2. Supported Hardware

---

Have another question? Just tag @inkeep.

[Hedda]

@BUG1989 spotted the hardware specification for this on cnx-soft blog and have three questions with some follow-up questions/links:

• 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/

1. Can you consider making a less expensive low-power "lite" variant that does not need active cooling and could be sold for a little less?

For some Edge AI use cases it is not good to require active cooling, so good idea to sell a low-temperature with just a larger heatsink.

3. Can it be used for real-time AI accelerated object detection as well as for video decoding and video enccoding (i.e. video transcoding)?

That is, can it be used to accelerate all type of video decode and video encode tasks for Frigate NVR too?

4. Any chance you be willing to also contribute code for this to the Home Assistant Operating System (HAOS Linux distrobution)?

• https://github.com/home-assistant/operating-system

Home Assistant OS in turn can run Frigate add-ons

• https://github.com/blakeblackshear/frigate-hass-addons

See these Home Assistant specific requests and discussions asking about support for this new M5Stack LLM-8850 AI accelerator adapter :

• HAOS support for M5Stack LLM-8850 M.2 Axera AX8850 AI accelerator? home-assistant/operating-system#4333

• https://community.home-assistant.io/t/m5stack-llm-8850-8gb-m-2-axera-ax8850-24-tops-ai-accelerator/936599

However the AI-accelerator could be useful for many things in HA OS, such as STT (Whisper), TTS (Piper), object recognition, and local LLMs:

• https://www.home-assistant.io/blog/2025/09/11/ai-in-home-assistant/

PS: Home Assistant Operating System supports 64-bit ARM and 64-bit x84 platforms but the most popular today is probably Raspberry Pi 5:

• https://en.wikipedia.org/wiki/Home_Assistant

[Hedda]

1. Can you consider making a less expensive low-power "lite" variant that does not need active cooling and could be sold for a little less?

@BUG1989 Oh, and would you also consider making a high-cost 16GB variant as well (for loading multiple models that don't fit 8GB)?

[MarsTechHAN]

It's not physically possible, as the maximum capacity of this DRAM controller is 8GB. I think 8GB is sufficient, though. The largest model it can run smoothly is around 1.5B Q4. A single language model only occupies about 700MB of memory space, leaving 6.3GB available. Not to mention that most vision models are only around 100MB in size.

[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/

Radxa has not yet posted anything about pricing or when it will be availability.

• https://linuxgizmos.com/radxa-launches-m-2-ai-accelerator-with-axera-ax8850-and-24-tops-npu/

Radxa officially lists their ROCK 5A, 5B, 5B+, and ROCK 5 ITX boards as tested.

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