Recipe by Saiyam Pathak Kubesimplify · based on Ollama, Qwen, NVIDIA
License: Blog post, no license stated · Ollama: MIT · Qwen: Apache-2.0.
TESTED BY Saiyam Pathak on 1x DGX Spark (GB10, 128 GB), Ollama v0.32.12, Q4_K_M · their results (benchmark table)
Qwen3.8-27B on one DGX Spark with Ollama
EASY: one DGX Spark, one command, 18 GB download.
FOROne NVIDIA DGX Spark.
NEEDOllama 0.32.12 or newer. Older versions refuse the download.
WHYFast for its size: 26.5 tok/s, because Ollama turns on the model’s own speed-up (MTP) by default.
INSTALL
curl -fsSL https://ollama.com/install.sh | shRUN
ollama run qwen3.8:27bCHECKollama ps. It should say 100% GPU. The author once saw it load on the CPU at 7.5 tok/s.
SAFETYOllama answers on this Spark only (port 11434) unless you change OLLAMA_HOST.
OR · LET A CODING AGENT DO IT
PASTEInto Claude Code, Codex or Cursor:
On my NVIDIA DGX Spark, install or update Ollama to 0.32.12 or newer with NVIDIA’s line (curl -fsSL https://ollama.com/install.sh | sh), then pull and start qwen3.8:27b.
Check with ollama ps that it runs 100% on the GPU and tell me the result.
Ask me before changing any system settings.If it fails
“requires a newer version” when you pull.
Update Ollama to 0.32.12 or newer.
About 7.5 tok/s.
It loaded on the CPU. Check ollama ps says 100% GPU; restart Ollama if not.
From the original README, shortened. Read it for full details.
What they tested
- Decode 26.5 tok/s (200 tokens), prefill 731 tok/s on a ~2,200-token prompt; temperature 0, 3 runs · evidence
- Ollama defaults to a 262K context on this box and wires up vision · evidence
Not shown in their results: No long-document test. One user only; no multi-user numbers for Ollama.
RunCards does not test recipes. We only list recipes whose authors publish their own test results. Checked 2026-10-19.