Qwen3.8 on DGX Spark (GB10)

This project installs a boot-persistent, hardened serving stack for the Qwen3.8 model family on a single DGX Spark (GB10), offering seven switchable serving targets, from the default NVFP4-quantized 27B model to an FP8 reference build, abliterated variants of each, and a 176B hybrid mixture-of-experts model with vision support, all served through SGLang. It reports measured throughput for each target, for example a greedy median tokens-per-second figure for the default 27B configuration using SGLang with NVFP4 quantization and speculative decoding, and a separate throughput and context-length figure for the 176B model.

The stack claims zero quality loss across every speculative decoding path by construction, and every switchable target is measured and compared for speed against the others, including comparisons of aggregate throughput at several concurrent stream counts. It ships as a single install command and a local cockpit for driving the resulting server.