Can Llama 4 Scout 17B 16E Instruct run on the DGX Spark?
Yes. At 8K context the best fit is Q6_K, using about 84.6 GB of the 109 GB usable on the DGX Spark. Estimated speed ceiling: ~19 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 38.4 GB Fits | 39.1 GB Fits | 43.6 GB Fits | 61.6 GB Fits | 42 |
| 3-bit (Q3_K_M) • | 49.7 GB Fits | 50.5 GB Fits | 55.0 GB Fits | 73.0 GB Fits | 32 |
| 4-bit (Q4_K_M) • | 62.4 GB Fits | 63.2 GB Fits | 67.7 GB Fits | 85.7 GB Fits | 25 |
| 5-bit (Q5_K_M) • | 72.8 GB Fits | 73.6 GB Fits | 78.1 GB Fits | 96.1 GB Fits | 21 |
| 6-bit (Q6_K) • | 83.9 GB Fits | 84.6 GB Fits | 89.1 GB Fits | 107 GB Tight | 19 |
| 8-bit (Q8_0) • | 108 GB Tight | 109 GB No | 113 GB No | 131 GB No | — |
| 16-bit (F16/BF16) • | 202 GB No | 203 GB No | 208 GB No | 226 GB No | — |
• real GGUF file size from unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF. Speed is a bandwidth ceiling (273 GB/s), not a benchmark. Methodology.
More
Llama 4 Scout 17B 16E Instruct on every GPU and Mac · Every model that runs on the DGX Spark · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (meta-llama/Llama-4-Scout-17B-16E-Instruct), NVIDIA specs. Data as of 2026-11-17.