DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s
| GPU(s) | Nvidia GB10 (unified) |
|---|---|
| VRAM / unified memory (GB) | 128 |
| CPU | Nvidia GB10 Grace (Arm) Amazon · eBay (used) · Newegg |
| Motherboard | |
| RAM | 128 GB LPDDR5X unified (~121 GB usable) |
| PSU | |
| Other | NVIDIA DGX Spark, DGX OS (aarch64), FP8 KV cache, 128K context per conversation |
| Reported total cost | |
| Power |
Measured speed (generation tokens/sec)
📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
| Model | Quant | Engine | tok/s | Details |
|---|---|---|---|---|
| Qwen3.6 35B-A3B | NVFP4 + MTP (k=2) | vLLM 0.31.0, 128K ctx, 1 user | 102.3 | ⚡ spec-decode (MTP) |
| Qwen3.6 35B-A3B | NVFP4, no speculative decoding | vLLM 0.31.0, 128K ctx, 1 user | 78.1 | |
| gpt-oss 20B | vLLM 0.31.0, 128K ctx, 1 user | 96.2 | ||
| Nemotron 3.5 Lightning 30B-A3B | + DSpark speculative decoding (k=3) | vLLM 0.31.0, 128K ctx, 1 user | 94.4 | ⚡ spec-decode (DSpark) |
| Nemotron 3.5 Lightning 30B-A3B | no speculative decoding | vLLM 0.31.0, 128K ctx, 1 user | 70.3 | |
| Nemotron 3 Nano Omni 30B | vLLM 0.31.0, 128K ctx, 1 user | 58.8 | ||
| Gemma 4 26B-A4B | + official MTP drafter (k=3) | vLLM 0.31.0, 128K ctx, 1 user | 56.5 | ⚡ spec-decode (MTP) |
| Gemma 4 26B-A4B | no speculative decoding | vLLM 0.31.0, 128K ctx, 1 user | 30.8 | |
| gpt-oss 120B | vLLM 0.31.0, 128K ctx, 1 user | 44.4 | ||
| Qwen3.8 Flash-Next | + MTP (k=2) | patched vLLM (qwen38-flash-dgx), 128K ctx, 1 user | 26.8 | ⚡ spec-decode (MTP) |
| Nemotron 3 Super 120B-A12B | + MTP (k=3) | vLLM 0.31.0, 128K ctx, 1 user | 23.0 | ⚡ spec-decode (MTP) |
| Nemotron 3 Super 120B-A12B | no speculative decoding | vLLM 0.31.0, 128K ctx, 1 user | 16.1 | |
| Qwen3.8 27B | + MTP (k=2) | vLLM 0.31.0, 128K ctx, 1 user | 19.0 | ⚡ spec-decode (MTP) |
| Qwen3.8 27B | no speculative decoding | vLLM 0.31.0, 128K ctx, 1 user | 11.0 | |
| Llama 3.1 8B | NVIDIA NIM 2.0.9, 128K ctx, 1 user | 26.4 |
Cristian Tala, last updated 6 Oct 2026. Single-user decode tok/s: median of 3 streamed requests with a ~200-word prompt; every model loaded with 128K context and memory budgeted for 4 concurrent users. Quant only stated for Qwen3.6 35B-A3B (NVFP4). With 4 users, total throughput was higher (e.g. 240 tok/s for Qwen3.6 35B-A3B + MTP). Eagle3 slowed gpt-oss down (96.2 to 59.4 on 20B), so it is not listed.
Source: https://ctala.github.io/local-llm-agentic-workflows/benchmarks/
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Quick answers: Can DGX Spark (GB10) run Gemma 4 26B-A4B? · Can DGX Spark (GB10) run Nemotron 3 Nano Omni 30B? · Can DGX Spark (GB10) run Nemotron 3 Super 120B-A12B? · Can DGX Spark (GB10) run Nemotron 3.5 Lightning 30B-A3B? · Can DGX Spark (GB10) run Qwen3.6 35B-A3B? · Can DGX Spark (GB10) run Qwen3.8 27B? · Can DGX Spark (GB10) run Qwen3.8 Flash-Next? · Can DGX Spark (GB10) run gpt-oss 120B? · Can DGX Spark (GB10) run gpt-oss 20B?