DGX Spark: every recipe for current models, measured

19 DGX Spark setups for 11 open models, from one Spark to four. Each number comes from a published, open-source recipe by Mia's AI Lab and links back to it.

Recipes table

Decode = single-stream generation speed. Prefill = prompt processing on a 32k prompt where Mia lists it. Sorted by number of Sparks.

Speculative decoding (MTP, DFlash, DSpark, LiLiCorr) is on in most rows. Engine names are as given on each recipe card; versions, patches and launch flags are in the linked recipe.

Memory math: how many Sparks does a model need?

Each Spark has 128 GB of unified memory (NVIDIA). The weights take about parameters × bits per weight ÷ 8, and the KV cache for long context needs room on top. NVFP4 stores 4-bit values with one FP8 scale per 16 values, so about 4.5 bits per weight (NVIDIA).

DeepSeek V4.1 Flash: Per the vLLM recipe, 196.6B of the parameters are Engram lookup tables (about 183 GiB) that vLLM can keep in CPU memory, so the GPUs mainly need the 552B backbone. The native checkpoint mixes MXFP4 experts with FP8 weights. Counting all 748.6B would need ≈271 GB, more than two Sparks.

Other DGX Spark builds

• Dell Pro Max with GB10 (Nvidia DGX Spark)
• DGX Spark (GB10, 128GB): Nemotron 3.5 Super VL 120B and Bolt 74B, launch-week runs
• DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s
• DGX Spark (GB10): Qwen3.6 35B-A3B NVFP4 at 113 tok/s on vLLM with MTP
• 4x DGX Spark cluster running GLM-5.2 (unpruned)
• 4x DGX Spark ring running GLM-5.3 Flash NVFP4
• 2x DGX Spark running GLM-5.3 Flash NVFP4
• DGX Spark with Qwen3.6 27B NVFP4 + DFlash speculative decoding
• 4x DGX Spark (GB10) running Step 5 Preview with tensor parallel over RoCE

All Spark results, including plain decode without speculation: DGX Spark (GB10) hub. Against other boxes: DGX Spark vs Mac Studio vs Strix Halo.

FAQ

How fast does GLM-5.3 Flash run on DGX Spark?

GLM-5.3 Flash does not fit one Spark at 4-bit (≈160 GB of weights). Mia's TensorFold recipe measures 60.4 tok/s on 2 Sparks (EXL3 4-bit + DFlash2, TensorFold; Mia's recipe), rising to 95.2 tok/s on 4 Sparks (EXL3 4-bit + DFlash2, TensorFold; Mia's recipe). These are single-stream decode speeds with speculative decoding.

What is the fastest current model on a single DGX Spark?

Among Mia's single-Spark recipes, Qwen3.8 Flash-Next reaches 48.7 tok/s on 1 Spark (NVFP4, vLLM; Mia's recipe). The highest single-Spark number in the table is 95.1 tok/s on 1 Spark (NVFP4 + MTP, vLLM; Mia's recipe), on a much smaller model.

Can DeepSeek V4.1 Flash run on DGX Sparks?

Yes, on two or more. The fastest single-stream figure we have is 74.9 tok/s on 2 Sparks (1017 checkpoint, EXL3 2.9-bit + DSpark, TensorFold Zig; Mia's recipe). The full model has 748.6B parameters, but per the vLLM recipe about 196.6B of them are Engram lookup tables that do not need to sit in accelerator memory.

How many DGX Sparks do you need for full GLM-5.3?

Three in Mia's recipe: 753B parameters at 2.75 bits is ≈259 GB of weights, more than two Sparks' 256 GB. Measured: 31.3 tok/s on 3 Sparks (EXL3 2.75-bit + DSpark, TensorFold; Mia's recipe).

How much memory does a DGX Spark have?

128 GB of unified LPDDR5X memory shared by the Grace CPU and the GB10 GPU (NVIDIA). Linking Sparks pools it: two give 256 GB, four give 512 GB, which is how the larger models in this table run.

Where do these numbers come from?

From Mia's AI Lab, which publishes open-source DGX Spark recipes (install scripts, engine patches and benchmarks) on GitHub. RigBase reads the recipe cards, checks each decode figure against the recipe page or README, and links every row back. We did not re-run them. Figures we could not find again in the recipe text are left out.

Are these numbers comparable with other GPUs?

Mostly not 1:1. Nearly every recipe uses speculative decoding (MTP, DFlash, DSpark, LiLiCorr), which speeds up predictable text far more than random text. Compare like with like, or see the plain-decode results on the DGX Spark hub.