Can NVIDIA DGX Spark (GB10) · 128 GB run Llama 4 Scout (109B-A17B)?

128 GB unified · 273 GB/s memory bandwidth · Llama 4 Scout (109B-A17B) has 108.6B parameters (17B active per token).

Short answer

Yes, comfortable for chat and acceptable or better for every workload we test.

Best quantization for chat: Q5_K_M (73 GB of weights) on Ollama. Generation: 12–16 tok/s, leaving about 45 GB free for your other apps.

How it rates for each workload

Same model and machine, different jobs. Agents and long documents send far bigger prompts than chat, so they need much more speed and memory.

Workload	Verdict	Quant	Generation	First reply	Context
Chatwith Open WebUI	Comfortable	Q5_K_M	12–16 tok/s	≈2 s	8K	Details
Coding questionswith Continue	Acceptable	Q5_K_M	11–15 tok/s	≈8 s	16K	Details
Coding in a repositorywith Aider	Acceptable	Q5_K_M	10–13 tok/s	≈30 s	32K	Details
Agentic codingwith OpenCode	Acceptable	Q4_K_M	10–14 tok/s	≈40 s	32K	Details
Long documents (64K)with Open WebUI	Acceptable	Q6_K	6–8 tok/s	≈2 min	64K	Details

Speeds are estimates calibrated against measured benchmarks on similar hardware. “First reply” is the time to read the workload’s opening prompt. How we calculate

Get it running

Copy-paste commands for the Q6_K build with a 64K context. Ollama and llama.cpp download the file straight from Hugging Face.

ollama run hf.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q6_K

This build is split into 2 files, which Ollama can't import from Hugging Face. Use llama.cpp or LM Studio instead.

Other models for NVIDIA DGX Spark (GB10) · 128 GB

The best-rated models for chat on this machine.

Everything NVIDIA DGX Spark (GB10) · 128 GB can run

Cheapest machines for Llama 4 Scout (109B-A17B)

Hardware rated comfortable or better for chat, lowest price first.

Same chip, other memory: 64 GB

Different tool, context size or workload?

These pages use typical settings. The full check models your exact setup.

Open weights (restricted)Check my exact setup