[Question] Qwen3-Omni Full Support on Thor/Orin (v0.6.0)

[Evan-Wang-26]

Hi Team,

Thanks for the v0.6.0 release

Hardware: NVIDIA Thor & Jetson Orin

Could you please clarify the current status of Qwen3-Omni:

1. Is the full multimodal pipeline (not just ASR/TTS) officially supported now?
2. Are there specific optimizations or performance benchmarks available for Thor and Orin?

Looking forward to your guidance. Thanks!

[fans-nv]

The work is under work. Because the 30B model is large, do you have a targeted quantized model to run on Orin?

[Evan-Wang-26]

Hi @fans-nv,

Thanks for the update.

Regarding the model size, we identified community quantized versions that should fit Orin 32GB:

1. GPTQ 4-bit: https://huggingface.co/thomasip/Qwen3-Omni-30B-A3B-Instruct-GPTQ-4bit
2. AWQ 4-bit: https://huggingface.co/cyankiwi/Qwen3-Omni-30B-A3B-Instruct-AWQ-4bit

Since our deployment strategy targets TRT-Edge-LLM, could you clarify:

1. Does TRT-Edge-LLM currently support loading such quantized MoE models directly?
2. Is there an ETA for official Qwen3-Omni integration?

We prefer to proceed with TRT-Edge-LLM if the compatibility is confirmed. Thanks!