What can a DGX Spark 128GB run?
128GB · 273 GB/s · 100W · $3999
NVIDIA's desk-sized mini-supercomputer — 128 GB of Grace-Blackwell memory for big MoE models at 100 W.
Runs great22
Comfortable headroom — daily-driver territory.
| Model | Quant | Est. speed | Max context |
|---|---|---|---|
| deepseek moe 16b chat | FP16 | Infinity tok/s | — |
| qwen2 7b instruct | FP16 | Infinity tok/s | — |
| Qwen3-0.6B | FP16 | 136 tok/s | 33k |
| Gemma-3n-E4B | FP16 | 60 tok/s | 131k |
| GPT-OSS-120B | Q4_K_M | 55 tok/s | 103k |
| SmolLM2-1.7B | FP16 | 50 tok/s | 8k |
| Llama-3.2-3B | FP16 | 26 tok/s | 131k |
| Qwen3-30B-A3B | FP16 | 26 tok/s | 33k |
| Qwen3-Coder-30B-A3B | FP16 | 26 tok/s | 195k |
| GPT-OSS-20B | FP16 | 24 tok/s | 131k |
| Phi-4-mini | FP16 | 22 tok/s | 131k |
| Gemma-3-4B | FP16 | 22 tok/s | 131k |
| Mistral-7B-Instruct | Q8_0 | 22 tok/s | 33k |
| Qwen3-4B | FP16 | 22 tok/s | 33k |
| DeepSeek-R1-Distill-Qwen-7B | Q8_0 | 21 tok/s | 131k |
| Qwen2.5-Coder-7B | Q8_0 | 21 tok/s | 131k |
| Llama-3.1-8B | Q8_0 | 20 tok/s | 131k |
| Qwen3-8B | Q8_0 | 20 tok/s | 33k |
| Qwen3-14B | Q4_K_M | 17 tok/s | 33k |
| Phi-4-14B | Q5_K_M | 16 tok/s | 16k |
| DeepSeek-R1-Distill-Qwen-14B | Q5_K_M | 16 tok/s | 131k |
| Gemma-3-12B | Q6_K | 16 tok/s | 131k |
Tight squeeze14
It fits, but close the browser tabs and mind the context.
| Model | Quant | Est. speed | Max context |
|---|---|---|---|
| Qwen3.8 Flash Next | Q3_K_M | 116 tok/s | 24k |
| Qwen3-235B-A22B | IQ2_M | 23 tok/s | 33k |
| Codestral-22B | FP16 | 4 tok/s | 33k |
| Mistral-Small-24B | FP16 | 4 tok/s | 131k |
| Gemma-3-27B | FP16 | 3 tok/s | 131k |
| qwen35 27B | FP16 | 3 tok/s | 241k |
| Qwen3.8 | FP16 | 3 tok/s | 246k |
| Qwen3.8 27B | FP16 | 3 tok/s | 246k |
| DeepSeek-R1-Distill-Qwen-32B | FP16 | 3 tok/s | 131k |
| GLM-4-32B | FP16 | 3 tok/s | 131k |
| Qwen3-32B | FP16 | 3 tok/s | 33k |
| Llama-3.3-70B | Q8_0 | 2 tok/s | 89k |
| Qwen2.5-72B | Q8_0 | 2 tok/s | 77k |
| glm-4.6-air 106B | Q6_K | 2 tok/s | 20k |
Measured on this card1
Everything here was benchmarked on this exact card by someone who owns one — the tables above are our arithmetic; these numbers came off real machines.
Deepseek V4 Flash 0731: fastest recorded generation is 60 tok/s — NVFP4 on 2× DGX Spark 128GB in vLLM.