Recipe by Vaishak Salin · based on Ollama, Google, NVIDIA
License: Blog post, no license stated · Ollama: MIT · Google: Apache-2.0.
TESTED BY Vaishak Salin on 1x DGX Spark (GB10, 128 GB), Ollama, Q4_K_M · their results (benchmark table)
Gemma 4 31B on one DGX Spark with Ollama
EASY: one DGX Spark, one command, 20 GB download.
curl -fsSL https://ollama.com/install.sh | shollama run gemma4:31bOR · LET A CODING AGENT DO IT
On my NVIDIA DGX Spark, install Ollama with NVIDIA’s line (curl -fsSL https://ollama.com/install.sh | sh) if it is missing, then pull and start gemma4:31b.
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
Replies are slow, about 9 tok/s.
Expected for this dense 31B model on a Spark. In the same test Gemma 4 26B ran at 59.4 tok/s: ollama run gemma4:26b.
From the original README, shortened. Read it for full details.
What they tested
- Gemma 4 31B, Ollama Q4_K_M, one user: 9.4 tok/s, 106 ms between tokens · evidence
- Gemma 4 26B, Ollama Q4_K_M, one user: 59.4 tok/s (same test) · evidence
Not shown in their results: Five requests per run; the author calls the numbers directional. Tested April 12, 2026 with the Ollama of that time. The post names Ollama Q4_K_M, not the exact tag. No long-document or image test.
RunCards does not test recipes. We only list recipes whose authors publish their own test results. Checked 2026-10-19.