devjeonghwan/vllm-for-rtx-spark is a repository that documents how to build vLLM and run Qwen3.8-Flash-Next-FP4-Local on an NVIDIA RTX Spark with Windows ARM64. The latest release is v0.1.1. The provided material contains no changelog for this version, so what changed in this release cannot be confirmed.
The configuration covered by the repository targets an environment with an RTX Spark N1X, 96 GB of graphics memory, 32 GB of CPU memory, and NVMe storage. It starts from the Qwen3.8-Flash-Next recipe in another repository and vLLM 0.32.0. It also walks through the installation and execution steps, the per-memory configuration files, and how to build the vLLM components directly. The README includes a performance comparison table against DGX Spark and llama.cpp, and the llama.cpp figures based on the same recipe were published by NVIDIA. The table compares code decode throughput and total generation time.
The performance table is therefore best read as information about that recipe and those comparison items. No material showing performance or applicability in other configurations is presented in FACTS. The specific changes in v0.1.1 are also not given, which makes it hard to judge new features of the release itself. The procedure is useful as a reference if you want to build vLLM on Windows ARM64, but there is not enough basis to apply the performance results as they are to other environments.