[Feature Request] Official NPU Support for Qwen3.6-35B-A3B MoE (.q4nx)

Hey team,

I'm looking to run local MoE-based text generation workflows directly on the NPU (currently deploying on a Strix Halo APU) and would heavily prefer to leverage Qwen3.6-35B-A3B for its excellent balance of inference speed and reasoning capability.

I see that the core Qwen 3 and Qwen 3.5 architectures are already mapped and supported for text generation and vision models in the registry. However, I noticed there isn't an officially published .q4nx file for the Qwen3.6-35B-A3B MoE variant yet.

I have a couple of questions regarding the path forward:

• Official Support: Are there plans to officially publish and maintain the .q4nx for this specific MoE model in the registry so it can be natively pulled via flm pull?

• Custom Compilation: If I were to compile it myself, are there any known gotchas for this specific MoE architecture? I'm particularly concerned about the routing logic for the 35B total / 3B active parameters and how the NPU sharding handles the expert layers versus the shared attention layers.

If there is a reference script or specific flm-build configuration you recommend for adapting the Qwen 3.5 MoE graph to handle this 3.6 variant, that would be incredibly helpful.

Thanks for the great work on the engine and the NPU acceleration stack!

Start by reviewing the existing Qwen 3 and Qwen 3.5 entries in the model registry, then inspect the Qwen 3.5 MoE graph and the flm-build path mentioned in the issue. Determine whether the Qwen3.6-35B-A3B architecture can be represented for .q4nx generation, including expert routing and NPU sharding, and define what registry support and successful flm pull behavior would confirm completion.