NVIDIA DGX Spark with Ollama: official performance numbers
Ollama tested the DGX Spark (GB10, 128GB unified memory) on release firmware: gpt-oss 20B at ~58 tok/s and gpt-oss 120B at ~41 tok/s decode, while dense 70B q4 managed ~4.4 tok/s.
Solo2-10 peopleCodingWriting assistantDocument search / chat with files$3k+Sensitivity: High (client/confidential)
Hardware
| Item | Approx. price (as of 2026-10-16) |
|---|---|
| On sale −6% MSI GeForce RTX 5060 Ti 16GB Ventus 3X OC (Newegg, related current offer) | $579.99 |
| NVIDIA DGX Spark (GB10 Grace Blackwell, 128GB unified memory) | Not stated in source |
Source prices are as stated by the source. Retailer offers: prices checked 2026-10-16, may change. Some links may be affiliate links; we may earn a commission at no extra cost to you. This never changes what we list.
Software stack
- Ollama v0.12.6
- Codex (with --oss)
Models used
- gpt-oss 20B / 120B (MXFP4)
- gemma3 12B/27B
- llama3.1 8B/70B
- qwen3 32B
- deepseek-r1 14B
Good at
- gpt-oss 20B: 58.27 tok/s decode; gpt-oss 120B: 41.14 tok/s
- Fits gpt-oss-120b entirely in memory
Weak at
- Dense models are slow: gemma3 27B q4_K_M 10.83 tok/s, 70B q4_K_M 4.423 tok/s
Rough total cost
Not stated in source.
Getting started
- Update DGX Spark firmware.
- Install Ollama.
- ollama run gpt-oss (or gpt-oss:120b).
- Optionally: codex --oss --model gpt-oss for local coding.
Source & credit
NVIDIA DGX Spark performance — Ollama blog
Summarized from the source; quotes are the source author's words. Check the original before buying.