NVIDIA’s DGX Spark (GB10) started as a deskside curiosity — a 128GB unified-memory workstation drawing ~38W from the GPU during inference. Over the last several weeks, a wave of open-source recipes and community benchmarks has turned the Spark into a modular building block. Users are connecting 2, 3, and 4 units directly — no switch required — and publishing reproducible configurations for models ranging from 35B parameters past 1 trillion.
We collected the published benchmarks across all four cluster sizes to understand where each configuration makes sense.
Single GB10 (128GB unified)
The single-Spark sweet spot is dense models under 35B parameters and MoE models with small active footprints:
| Model | Total | Active | Quant | Speed | Quality | Source |
|---|---|---|---|---|---|---|
| Qwen3.6-35B-A3B | 35B | 3B (A3B) | NVFP4 | 700+ tok/s agg (64 users) | Excellent | spark-bench |
| Qwen3.6-27B-NVFP4 | 27B | 27B (dense) | NVFP4 | 84.7→183 tok/s (b16) | Very high | NVIDIA on HF |
| Nemotron Puzzle 75B-A9B | 75B | 9.3B (MoE+Mamba) | NVFP4 | ~38-42 solo / ~75 agg | Strong | GitHub recipe |
| Nemotron-3-Super-120B-A12B | 120B | 12B (MoE) | NVFP4 | ~20-30 tok/s | Frontier-ish | community post |
| Hy3 (Tencent) | 295B | ~21B (MoE, 8/192) | FP8/NVFP4 | ~30 tok/s (128GB Mac) | Beats DS V4 | Tencent |
| DeepSeek V4 Flash (GGUF) | ~671B | ~37B (MoE) | 8-bit GGUF | ~8-15 tok/s | Frontier | Unsloth |
| GLM-5.2 (Colibri, 25GB RAM) | 744B | ~40B (MoE) | Colibri C (disk) | ~4-8 tok/s | Works on laptop | independent dev |
We think the single Spark makes the most sense for organizations that need to serve multiple models side by side, or run inference and fine-tuning on separate nodes. At roughly $4K per unit, there is no other Blackwell-class compute at this power envelope.
Paired GB10s (256GB unified, approximately $8K total)
With two Sparks connected directly, the range of models expands meaningfully:
| Model | Total | Active | Quant | Speed | Quality | Source |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash | ~671B | ~37B (MoE) | NVFP4 | 55-60 tok/s, 1M ctx | Frontier | community post |
| DeepSeek V4 Flash (NVFP4 KV) | ~671B | ~37B (MoE) | NVFP4 KV cache | ~55-60 tok/s, 1.5M pool | Frontier | NVFP4 KV recipe |
| GLM-5.2 744B (pruned) | 744B | ~40B (MoE) | NVFP4 + 15% prune | ~15→24 tok/s (DPC4+MTP4) | Near-full acc, 256K | community post |
| MiniMax M3.0 | 428B | ~45B (MoE) | NVFP4 | 36 tok/s | High | vLLM PR |
We think this is the best price-to-performance point in the local-inference market today. The 256GB configuration can run DeepSeek V4 Flash at 55-60 tok/s with a 1M-token context for $8,000 — roughly the cost of a single NVIDIA RTX PRO 6000 with less memory and no networking. The paired configuration hits a performance ceiling around 300B unquantized parameters; beyond that, quantization becomes necessary.
Three GB10s (384GB unified, approximately $12K)
The three-Spark cluster is the most thoroughly documented configuration in the community, largely because of the MiMo V2.5 Omni work:
| Model | Total | Active | Quant | Speed | Quality | Source |
|---|---|---|---|---|---|---|
| MiMo V2.5 Omni | 310B | 15B (MoE) | FP8/NVFP4 | ~39 tok/s | 97.3 TrueScore | full recipe |
| MiMo V2.5 (concurrency) | 310B | 15B (MoE) | FP8/NVFP4 | 37→82 tok/s (1→8 agents) | 94.4 TrueScore | benchmark |
| MiMo V2.5 (thinking ON/OFF) | 310B | 15B (MoE) | FP8/NVFP4 | 97.3@1.2s / 88.9@2.4s | Detailed eval | eval results |
| GLM-5.2 744B (no pruning) | 744B | ~40B (MoE) | NVFP4+FP3+MXFP8 | Full vLLM speed | Near-full acc | vLLM config |
| GLM-5.2 REAP 469B | 469B | ~40B (MoE) | NVFP4 | ~4.4 tok/s, 256K ctx | Good | REAP post |
| GLM-5.2 vs Kimi K2.7 (~1T) | 754B/~1T | ~40B/~40B (MoE) | 2-bit | 7.6 / 10.6 tok/s | 96.4 / 93.5 | comparison |
| MiniMax M3.0 | 428B | ~45B (MoE) | NVFP4 | ~36 tok/s | High | vLLM PR |
The three-Spark setup requires no external switch — the units connect directly. The 384GB of unified memory is the threshold where full 744B models like GLM-5.2 fit without pruning. MiMo V2.5 Omni is the standout: 97.3 TrueScore at 1.2 seconds with thinking off, full multimodal (text, image, video, audio), and 82 tok/s aggregate across eight agents.
Four GB10s (512GB unified, approximately $16K) and Beyond
| Model | Total | Active | Quant | Speed | Quality | Source |
|---|---|---|---|---|---|---|
| GLM-5.2 744B (NVFP4+15% prune) | 744B | ~40B (MoE) | NVFP4+FP3+MXFP8 | ~22 tok/s, MTP | Near-full acc | community post |
| GLM-5.2 744B (DPC4+MTP4) | 744B | ~40B (MoE) | NVFP4 | ~23-24 tok/s @ 128K | Near-full acc | optimization |
| GLM-5.2 744B (best config) | 744B | ~40B (MoE) | NVFP4+FP3 | Higher than 22 tok/s | Full, no prune | best config |
A community workaround exists for connecting four Sparks without a switch. At the extreme end, one builder assembled six GX10s (the ASUS-branded Spark), two Ultra nodes, and two Minis into a 1.5TB unified-memory cluster for $17,000, repurposing retired Ethereum mining frames.
Where Each Configuration Fits
| Configuration | Unified Memory | Cost | Leading Model | Performance | Best for |
|---|---|---|---|---|---|
| 1x GB10 | 128 GB | ~$4K | Qwen3.6-35B-A3B | 700+ tok/s (64 users) | Multi-model serving, single-model inference |
| 2x GB10 | 256 GB | ~$8K | DeepSeek V4 Flash | 55-60 tok/s, 1M ctx | Best price-to-performance for frontier models |
| 3x GB10 | 384 GB | ~$12K | MiMo V2.5 Omni | 39 tok/s, 97.3 TrueScore | Full multimodal, largest model selection |
| 4x GB10 | 512 GB | ~$16K | GLM-5.2 (full) | 22-24 tok/s | Maximum capacity per node, full-precision MoE |
We think the most interesting inflection point is at two Sparks. The 256GB configuration can run DeepSeek V4 Flash at 55-60 tok/s with a 1M-token context for $8,000 — roughly the cost of a single NVIDIA RTX PRO 6000 with less memory and no cabling or networking. The three-Spark cluster, at 384GB, is the threshold where full 744B models fit without pruning. Above that, the incremental memory matters less than the community recipe ecosystem, which is currently densest at the two- and three-Spark levels.