Qwen3-VL 8B

User reports Qwen3-VL-8B-Instruct at ~88-92 tok/s on an NVIDIA Jetson AGX Thor with the fast/tight 8K KV config, and ~70-78 tok/s with the 128K KV long-context config. Setup is a single Jetson AGX Thor with 64GB unified memory, TensorRT-Edge-LLM, NVFP4 weights and EAGLE-3 speculative decoding. EAGLE-3 averages ~3.76 accepted tokens per verify pass; cold start per request is ~6-8s and only one inference runs at a time.