Sorry if I’m late for this party but I just received my o6n ‘s (got two), and I was quite surprised to discover that the current workflow requires compiling models on an x86 host machine, rather than directly on ARM.
According to the official documentation, the NPU toolchain (model conversion / compilation) is only available for x86 systems, which effectively forces ARM developers to rely on a separate x86 machine in the development loop. (can’t even use an M based mac)
For an ARM-first platform, this is a significant limitation:
-
It breaks native, on-device development workflows.
-
It complicates CI/CD pipelines for ARM-based systems.
-
It makes headless, self-hosted, or fully ARM clusters impractical.
-
It adds friction for open-source and research use cases, where reproducibility and portability matter.
This is particularly disappointing given that, imo, many users interested in the Orion O6(N) are doing so specifically to build ARM-native AI systems, edge clusters, or low-power inference setups without depending on x86 infrastructure.
I would like to ask the Radxa team (or anyone with insight):
-
Are there any plans to port the NPU compiler / toolchain to ARM?
-
If not, is there any possibility of open-sourcing parts of the compiler toolchain so the community can attempt an ARM port?
-
Is this limitation due to licensing constraints, third-party IP, or purely engineering resources?
-
Is there at least a roadmap or timeline being considered for reducing this dependency?
Requiring an x86 machine to unlock a key feature of an ARM board feels at odds with the platform’s positioning and will likely be a blocker for many advanced users.
Any clarification would be appreciated, especially regarding future plans or community involvement options.
Thank you.