DGX Spark vs RTX 5090
Best sourced output tok/s per model (single-stream generation) — engine and version, quantization, context and a source link on every row. Hover a value for the quant.
DGX Spark48 sourced runs · best 86.8 tok/s
RTX 509023 verified runs · best 446.7 tok/s
Specs and evidence
| DGX Spark | RTX 5090 | |
|---|---|---|
| VRAM | 128 GB | 32 GB |
| Sourced runs on the board | 48 | 23 |
| Models measured | 9 | 16 |
| Best output tok/s | 86.8 tok/s | 446.7 tok/s |
| Highest measured peak VRAM | — | 31.3 GB |
| Street price (latest snapshot) | — | — |
Head-to-head
Only one model has been measured on both cards (Qwen3.8-27B). On that single data point DGX Spark records 56.0 tok/s and RTX 5090 records 435.1 tok/s — treat it as one data point, not a ranking.
The measured matrix
| Model | DGX Spark | RTX 5090 |
|---|---|---|
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B | — | 446.7 |
| Qwen3.8-27B | 56.0 | 435.1 |
| Ornith-1.0-35B-PrismaAURA-4.75bit-vllm-MTP | — | 283.1 |
| Kimi-Linear-48B-A3B-Instruct | — | 280.8 |
| Ornith-1.5-35B-A3B | — | 258.8 |
| North-Mini-Code-1.0 | — | 258.4 |
| Qwen3.6-35B-A3B-MTP | — | 222.9 |
| Laguna-XS-2.1 | — | 216.7 |
| KAT-Coder-V2.5-Dev-AWQ-W4A16-ASYM | — | 201.2 |
| Ornith-1.0-35B | — | 196.6 |
| Muse-Glimmer-30B | — | 124.8 |
| Qwen3.6-27B | — | 120.9 |
| Huihui-Qwen3.8-27B-abliterated | — | 117.7 |
| Qwen3.8-27B-5090-goldilocks | — | 93.5 |
| Qwen 3.8 Flash | 86.8 | — |
| Qwen3.8-Flash-Next | 75.7 | — |
| GLM-5.3-Flash | 64.1 | — |
| CYBER-FROST-3.8 | 60.7 | — |
| Qwen3.8-Flash-Next-hibrid48 | 60.0 | — |
| NVIDIA-Nemotron-3.5-Super-VL-120B-A12B | 48.0 | — |
| DeepSeek-V4.1-Flash | 47.0 | — |
| Qwen3.6-27B-MTP | — | 28.3 |
| DeepSeek-V4-Flash-0731 | — | 26.4 |
| Underdog-Saluki-27B | 22.6 | — |