Qwen3.8-Flash-Next on one box
180B parameters on a single Spark by keeping the 51B n-gram embedding table on NVMe instead of in memory. One install script, and a model that has no business fitting in 128 GB runs anyway.
Collection · one GB10 box, pushed hard
NVIDIA
DGX Spark
Everything I have published about serving large models on one machine. Every figure was recorded on the hardware.
Qwen3.8-27B FP8
From 7.88 to 58.5 tok/s single-stream with the weights untouched — the gain comes entirely from decode strategy. Speculative decoding and prefix caching benchmarked side by side, plus DFlash 2, the only Qwen3.8-27B build that serves under vLLM.
Qwen3.8-27B 4-bit
75 tok/s single-stream, 246 tok/s aggregate at 8-way concurrency. NVFP4 against MixedInt4-AutoRound against the FP8 baseline on one harness — including why the quantisation advantage collapses to +0.2% by c16.
DeepSeek V4 Flash
384k context, thinking, tools and an OpenAI-compatible API on one Spark, installed with a single command. The boring part done properly so the interesting part is reproducible.
Inference Atlas
The wider project these runs feed into: a community-owned map of inference engine configurations, where the repo is the database. DGX Spark is one column among many, and the grey squares are the point.