DGX Spark is NVIDIA's attempt to put a DGX on your desk. It's a small, fan-cooled box (about the size of a Mac mini) built around the GB10 Grace Blackwell superchip: a Blackwell GPU die plus a Grace CPU die, connected by NVLink-C2C, sharing 128GB of unified memory. It's the first NVIDIA machine that works like an Apple Silicon: one pool of memory, no PCIe transfers for model weights.
The numbers
- Memory: 128GB unified LPDDR5x at ~273GB/s. This is the constraint, not the compute.
- Compute: ~1000 TFLOPS FP4 (1 PFLOP), ~500 TFLOPS FP8, ~250 TFLOPS FP16. Blackwell tensor cores with NVFP4 support.
- CPU: 20-core Arm Grace CPU (10 performance + 10 efficiency).
- Power: 400W TDP, air-cooled, silent-ish.
- Price: $4,699 (128GB) / $3,999 (128GB, 1TB SSD config).
Measured, not marketing
local.ai has 84 measured benchmarks on the DGX Spark. The standout: it's the only personal machine that runs NVFP4 weights via vLLM at interactive speeds. Here's the top of the list:
- Qwen3.6 35B A3B (nvfp4, vLLM): 120.8 tok/s, τ² 98.0%
- LFM2.5 8B A1B (Q8_0): 120.0 tok/s, τ² 94.8%
- Laguna XS 2.1 (nvfp4): 103.0 tok/s, GAIA 56.2%
- Gemma 4 E2B It (UD-Q4_K_XL): 100.2 tok/s
- Diffusiongemma 26B A4B It (nvfp4): 96.8 tok/s, τ² 65.7%
- Nemotron 3.5 Lightning (nvfp4): 88.6 tok/s, GAIA 68.5%
- Qwen3.6 35B A3B (nvfp4): 86.3 tok/s, τ² 97.9%
- Qwen3.6 35B A3B (nvfp4): 82.9 tok/s, τ² 96.5%
- Laguna XS 2.1 (Q4_K_M): 78.6 tok/s, GAIA 64.0%
- Qwen3.6 35B A3B (iq1m): 78.4 tok/s, τ² 98.6%
The pattern is unmistakable: NVFP4 via vLLM is the way to run this machine. The same 35B A3B model gets 120.8 tok/s in NVFP4 (vLLM) vs 78.4 tok/s in IQ1M (llama.cpp), a 54% speedup from format alone. That's Blackwell's micro-tensor scaling doing real work.
Why it matters for inference
DGX Spark is the first personal machine that runs the same weight format and the same engine (vLLM, NVFP4) as a production cluster. That means the quantization you test at home is the quantization you ship. It's a 128GB unified-memory machine, so a 70B at 4-bit (~40GB) leaves room for KV cache and context. A 405B at 4-bit (~230GB) needs two, but a 671B MoE at 4-bit? Not on one.
It's not a B200. It's not meant to be. It's the machine that makes local frontier inference a hobby you can afford, and the benchmark data proves it's not a toy.
Sources
- local.ai hardware benchmarks: 84 measured runs on DGX Spark.
- NVIDIA DGX Spark product page: official specs.
- NVIDIA GB10 superchip details: the Grace Blackwell story.