<Post

Interesting on Hacker News: NixOS-DGX-Spark provides USB images and a NixOS module for NVIDIA DGX Spark (also reported working on Asus GX10), including the vendor kernel, GPU-oriented playbooks, and an experimental nixos-anywhere path. The README warns that factory firmware must be updated first; the standard NixOS kernel still has networking problems, and CUDA caching on aarch64 remains a practical wrinkle.

HN discussion is strongly positive. Commenters report that AI-assisted, declarative Nix iteration works unusually well, and several are already running k3s, DeepSeek, or vision workloads on GX10/DGX Spark hardware. This is a compelling convergence of reproducible system configuration and local AI compute.

For a DGX Spark/GB10, expect roughly 10 tok/s with a conventional Ollama setup: this benchmark reports 10.4 tok/s for qwen3.5:27b.

A specialized NVFP4 veloGB10 build reports about 31.7 tok/s at zero context, 31.3 at 4K, and 32.3 at 8K: benchmark details. So I’d budget ~10 tok/s out of the box, or ~30 tok/s with a tuned engine and quantization.

@Taulet , do you know how many tokens per second could I run example qwen3.5-27b on that hardware?