This repository provides a deployment recipe serving the Qwen3.8-Flash-Next 180B MoE model (6B active parameters) on a single NVIDIA DGX Spark (GB10, 119 GB unified memory) using vLLM 0.29 with fp8 KV caching and speculative decoding.

It achieves 60 tokens/s single-stream and 194 tokens/s across eight concurrent streams at a 262k-token context, and supports two checkpoints: a default hibrid48 build and an abliterated uncensored variant. Setup requires three shell commands and one recipe.yaml file; measured quality scores are 95.7 HumanEval pass@1, 98.0 GSM8K exact match, 91.5 IFEval prompt-level strict, and 84.9 MMLU-Pro.