What LLMs Can the NVIDIA DGX Spark 128GB Run?

The NVIDIA DGX Spark 128GB has 128 GB of unified memory (about 112 GB usable by the GPU) and 273 GB/s of bandwidth. It runs 51 of 58 models fully at Q4_K_M with 8K context, up to Qwen 3.5 122B-A10B, and generates about 30 tokens/s on an 8B model.

NVIDIA · Grace Blackwell · Released 2025-10 · Updated October 14, 2026

51 of 58 models fit

NVIDIA DGX Spark 128GB (112 GB GPU-usable) at Q4_K_M with 8K context. Largest that fits fully: Qwen 3.5 122B-A10B.

ModelNeedsStatusSpeedBest quant that fits
Qwen 3.5 122B-A10B70.2 GBFits~23 tok/sQ6_K
Mistral Small 4 119B68.2 GBFits~35 tok/sQ6_K
gpt-oss-120b67.2 GBFits~43 tok/sMXFP4 (official)
Llama 4 Scout64.0 GBFits~14 tok/sQ6_K
GLM-4.5-Air62.1 GBFits~19 tok/sQ6_K
Qwen3-Next 80B-A3B46.3 GBFits~59 tok/sQ8_0
Qwen 2.5 72B44.5 GBFits~3.6 tok/sQ8_0
Llama 3.3 70B43.2 GBFits~3.7 tok/sQ8_0
DeepSeek R1 Distill Llama 70B43.2 GBFits~3.7 tok/sQ8_0
Mixtral 8x7B28.0 GBFits~18 tok/sFP16 / BF16
Qwen 2.5 32B21.4 GBFits~7.8 tok/sFP16 / BF16
Qwen 2.5 Coder 32B21.4 GBFits~7.8 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 32B21.4 GBFits~7.8 tok/sFP16 / BF16
Qwen 3 32B21.4 GBFits~7.8 tok/sFP16 / BF16
Qwen 3.5 35B-A3B21.0 GBFits~62 tok/sFP16 / BF16
Qwen 3.6 35B-A3B21.0 GBFits~62 tok/sFP16 / BF16
Gemma 4 31B19.8 GBFits~8.2 tok/sFP16 / BF16
Qwen 3 30B-A3B18.8 GBFits~51 tok/sFP16 / BF16
Qwen3-Coder 30B-A3B18.8 GBFits~51 tok/sFP16 / BF16
Gemma 3 27B17.6 GBFits~9.4 tok/sFP16 / BF16
Qwen 3.5 27B16.9 GBFits~9.7 tok/sFP16 / BF16
Qwen 3.6 27B16.9 GBFits~9.7 tok/sFP16 / BF16
Qwen 3.8 27B16.9 GBFits~9.7 tok/sFP16 / BF16
Mistral Small 3.1 24B15.6 GBFits~11 tok/sFP16 / BF16
Gemma 4 26B-A4B15.6 GBFits~53 tok/sFP16 / BF16
gpt-oss-20b12.9 GBFits~61 tok/sMXFP4 (official)
Qwen 2.5 14B10.7 GBFits~17 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 14B10.7 GBFits~17 tok/sFP16 / BF16
Phi-4 14B10.6 GBFits~17 tok/sFP16 / BF16
Qwen 3 14B10.5 GBFits~17 tok/sFP16 / BF16
Ministral 3 14B9.7 GBFits~18 tok/sFP16 / BF16
Mistral Nemo 12B9.0 GBFits~20 tok/sFP16 / BF16
Gemma 3 12B8.8 GBFits~20 tok/sFP16 / BF16
Ministral 3 8B6.7 GBFits~28 tok/sFP16 / BF16
Qwen 3 8B6.6 GBFits~29 tok/sFP16 / BF16
Qwen 3.5 9B6.4 GBFits~29 tok/sFP16 / BF16
Llama 3.1 8B6.3 GBFits~30 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 8B6.3 GBFits~30 tok/sFP16 / BF16
Mistral 7B v0.35.7 GBFits~33 tok/sFP16 / BF16
Gemma 4 E4B5.7 GBFits~33 tok/sFP16 / BF16
Qwen 2.5 7B5.6 GBFits~33 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 7B5.6 GBFits~33 tok/sFP16 / BF16
Qwen 3 4B4.2 GBFits~53 tok/sFP16 / BF16
Phi-4 Mini4.1 GBFits~56 tok/sFP16 / BF16
Gemma 4 E2B4.0 GBFits~51 tok/sFP16 / BF16
Gemma 3 4B3.7 GBFits~57 tok/sFP16 / BF16
Qwen 3.5 4B3.7 GBFits~59 tok/sFP16 / BF16
Ministral 3 3B3.5 GBFits~64 tok/sFP16 / BF16
Llama 3.2 3B3.5 GBFits~67 tok/sFP16 / BF16
Llama 3.2 1B1.7 GBFits~180 tok/sFP16 / BF16
Gemma 3 1B1.6 GBFits~242 tok/sFP16 / BF16
Kimi K2.6566 GBToo small––
DeepSeek R1 671B380 GBToo small––
Llama 3.1 405B234 GBToo small––
Qwen 3.5 397B-A17B226 GBToo small––
DeepSeek V4 Flash161 GBToo small–Q2_K
Qwen 3 235B-A22B135 GBToo small–Q2_K
MiniMax-M2133 GBToo small–Q2_K

NVIDIA DGX Spark 128GB for local AI

A compact Grace Blackwell system with 128 GB of unified memory and the full CUDA stack. Like Strix Halo, its 273 GB/s bandwidth favours MoE models over large dense ones.

Setup tips: it runs NVIDIA's DGX OS with CUDA preinstalled, so vLLM, TensorRT-LLM and PyTorch work as on a datacenter GPU. Use it for MoE models and for fine-tuning experiments; dense 70B models run, but slowly, because of its 273 GB/s bandwidth.

Our picks for the NVIDIA DGX Spark 128GB

The model we would run first for each job, chosen from our recommendations by what fits fully with an 8K context. "Best quant" is the highest quality that fits; "max context" is at Q4_K_M or the native format.

TaskModelBest quantMax contextSpeed
Best all-rounderQwen 3.5 122B-A10BQ6_K256K~23 tok/s
CodingGLM-4.5-AirQ6_K128K~19 tok/s
Fastest repliesQwen 3.5 122B-A10BQ6_K256K~23 tok/s
Long documents (64K+)Qwen 3.5 122B-A10BQ6_K256K~23 tok/s
Images in, text outGemma 4 31BFP16 / BF16256K~8.2 tok/s
Image generationFLUX.2 [dev]FP8––
SpecValue
Memory128 GB unified (112 GB GPU-usable by default)
Memory bandwidth273 GB/s
ArchitectureGrace Blackwell
Released2025-10
Llama 3.1 8B Q4_K_M speed~30 tok/s
Largest model that fitsQwen 3.5 122B-A10B

Image and video models on the NVIDIA DGX Spark 128GB

Peak VRAM with model swapping (ComfyUI default on smaller cards).

ModelFP8FP8 fitGGUF Q4Q4 fit
Stable Diffusion 1.51.8 GBFits1.7 GBFits
SDXL 1.04.1 GBFits4.1 GBFits
Stable Diffusion 3.5 Medium6.4 GBFits6.4 GBFits
Stable Diffusion 3.5 Large8.6 GBFits6.4 GBFits
FLUX.1 [dev]12.1 GBFits7.3 GBFits
FLUX.1 [schnell]12.1 GBFits7.3 GBFits
Qwen-Image20.4 GBFits12.3 GBFits
Wan 2.2 A14B (T2V / I2V)18.8 GBFits13.2 GBFits
Wan 2.2 TI2V 5B8.0 GBFits6.5 GBFits
HunyuanVideo18.3 GBFits13.1 GBFits
FLUX.1 Kontext [dev]12.3 GBFits7.5 GBFits
FLUX.2 [dev]31.0 GBFits23.6 GBFits
Z-Image Turbo6.8 GBFits4.9 GBFits
Qwen-Image-Edit 250920.5 GBFits12.4 GBFits
LTX-219.2 GBFits12.6 GBFits

Comparisons

NVIDIA DGX Spark 128GB for LLMs: common questions

Can the NVIDIA DGX Spark 128GB run Llama 3.3 70B?

Yes. Llama 3.3 70B needs about 43.2 GB at Q4_K_M with 8K context, and the best quality that fits is Q8_0.

Can the NVIDIA DGX Spark 128GB run Qwen 3 32B?

Yes. Qwen 3 32B needs about 21.4 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

Can the NVIDIA DGX Spark 128GB run Qwen 3 14B?

Yes. Qwen 3 14B needs about 10.5 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

How fast is the NVIDIA DGX Spark 128GB for LLMs?

About 30 tokens per second for an 8B model at Q4_K_M, based on its 273 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.

Related