What can an NVIDIA DGX Spark (GB10) 128GB run?
The NVIDIA DGX Spark (GB10) 128GB has 112GB of unified memory and 273 GB/s of memory bandwidth. Run through the same engine as our calculator, it runs 38 of 51 current local LLMs fully on the GPU — the biggest is Qwen3 235B-A22B.
Every model, checked against the NVIDIA DGX Spark (GB10) 128GB
Best quant means the highest-quality quantization that still fits fully in 112GB at 8K context. Speeds are llama.cpp estimates from memory bandwidth — the same math the calculator uses.
| Model | Best quant | VRAM needed | Est. speed | Status |
|---|---|---|---|---|
| Llama 3.2 1B Llama | Q8_0 | 2.3 GB | ~146 tok/s | Runs on GPU |
| Qwen3 0.6B Qwen | Q8_0 | 1.7 GB | ~200 tok/s | Runs on GPU |
| Gemma 3 1B Gemma | Q8_0 | 2.0 GB | ~177 tok/s | Runs on GPU |
| Llama 3.2 3B Llama | Q8_0 | 5.1 GB | ~53 tok/s | Runs on GPU |
| Qwen3 1.7B Qwen | Q8_0 | 3.4 GB | ~85 tok/s | Runs on GPU |
| Phi-4-mini Phi | Q8_0 | 5.9 GB | ~45 tok/s | Runs on GPU |
| Qwen3 4B Qwen | Q8_0 | 5.8 GB | ~46 tok/s | Runs on GPU |
| Gemma 3 4B Gemma | Q8_0 | 6.8 GB | ~39 tok/s | Runs on GPU |
| Llama 3.1 8B Llama | Q8_0 | 10.6 GB | ~24 tok/s | Runs on GPU |
| Qwen3 8B Qwen | Q8_0 | 11.0 GB | ~23 tok/s | Runs on GPU |
| Mistral 7B v0.3 Mistral | Q8_0 | 9.8 GB | ~26 tok/s | Runs on GPU |
| Ministral 8B Mistral | Q8_0 | 10.7 GB | ~24 tok/s | Runs on GPU |
| gpt-oss 20B OpenAI | Q8_0 | 24.3 GB | ~56 tok/s | Runs on GPU |
| Qwen3 14B Qwen | Q8_0 | 18.5 GB | ~14 tok/s | Runs on GPU |
| Phi-4 14B Phi | Q8_0 | 18.7 GB | ~13 tok/s | Runs on GPU |
| Gemma 3 12B Gemma | Q8_0 | 17.2 GB | ~15 tok/s | Runs on GPU |
| Mistral Small 3.1 24B Mistral | Q8_0 | 29.1 GB | ~9 tok/s | Runs on GPU |
| Gemma 3 27B Gemma | Q8_0 | 36.6 GB | ~7 tok/s | Runs on GPU |
| Gemma 4 31B Gemma | Q8_0 | 40.1 GB | ~6 tok/s | Runs on GPU |
| Gemma 4 26B-A4B Gemma | Q8_0 | 30.1 GB | ~41 tok/s | Runs on GPU |
| Qwen3 32B Qwen | Q8_0 | 38.5 GB | ~6 tok/s | Runs on GPU |
| Qwen3 30B-A3B Qwen | Q8_0 | 35.0 GB | ~59 tok/s | Runs on GPU |
| Qwen3-Coder 30B-A3B Qwen | Q8_0 | 35.0 GB | ~59 tok/s | Runs on GPU |
| DeepSeek-R1 Distill 32B DeepSeek | Q8_0 | 39.3 GB | ~6 tok/s | Runs on GPU |
| Llama 3.3 70B Llama | Q8_0 | 82.0 GB | ~3 tok/s | Runs on GPU |
| Llama 4 Scout Llama | Q6_K | 96.1 GB | ~15 tok/s | Runs on GPU |
| gpt-oss 120B OpenAI | Q6_K | 101.8 GB | ~50 tok/s | Runs on GPU |
| Qwen3 235B-A22B Qwen | Q2_K | 104.7 GB | ~23 tok/s | Runs on GPU |
| Llama 4 Maverick Llama | — | 255.8 GB | — | Won't fit |
| DeepSeek-V3 DeepSeek | — | 427.8 GB | — | Won't fit |
| DeepSeek-R1 DeepSeek | — | 427.8 GB | — | Won't fit |
| Qwen3.5 9B Qwen | Q8_0 | 10.9 GB | ~24 tok/s | Runs on GPU |
| Qwen3.5 35B-A3B Qwen | Q8_0 | 39.7 GB | ~71 tok/s | Runs on GPU |
| Qwen3.6 35B-A3B Qwen | Q8_0 | 39.7 GB | ~71 tok/s | Runs on GPU |
| Qwen3.6 27B Qwen | Q8_0 | 31.2 GB | ~8 tok/s | Runs on GPU |
| Qwen3.8 27B Qwen | Q8_0 | 32.0 GB | ~8 tok/s | Runs on GPU |
| Gemma 4 E4B Gemma | Q8_0 | 10.2 GB | ~42 tok/s | Runs on GPU |
| DeepSeek V4-Flash DeepSeek | — | 181.4 GB | — | Won't fit |
| GLM-5.3-Flash GLM | — | 203.6 GB | — | Won't fit |
| Kimi K2.6 Kimi | — | 635.6 GB | — | Won't fit |
| MiniMax M3 MiniMax | — | 272.9 GB | — | Won't fit |
| Mistral Small 4 Mistral | Q6_K | 103.4 GB | ~41 tok/s | Runs on GPU |
| Kimi K3 Kimi | — | 1763.3 GB | — | Won't fit |
| Muse Spark Open Muse | — | 520.0 GB | — | Won't fit |
| Flint Recursive | Q8_0 | 45.2 GB | ~80 tok/s | Runs on GPU |
| Beam 1.0 Reflection | — | 318.7 GB | — | Won't fit |
| MiniMax M3.1 Flash MiniMax | — | 272.9 GB | — | Won't fit |
| Mistral Large 4 Mistral | — | 667.3 GB | — | Won't fit |
| Ling-3.1-flash-1019 Ling | — | 355.8 GB | — | Won't fit |
| LFM3-1.7B Liquid | Q8_0 | 2.6 GB | ~120 tok/s | Runs on GPU |
| LFM3-VL-2B Liquid | Q8_0 | 2.8 GB | ~108 tok/s | Runs on GPU |
NVIDIA DGX Spark (GB10) 128GB FAQ
What's the biggest LLM an NVIDIA DGX Spark (GB10) 128GB can run?
Fully on-GPU at 8K context, the biggest model that fits the NVIDIA DGX Spark (GB10) 128GB (112GB) is Qwen3 235B-A22B at Q2_K (~23 tok/s). Larger models need a card with more VRAM or system-RAM offload, which is much slower.
How fast is an NVIDIA DGX Spark (GB10) 128GB for local AI?
On a mid-size model like Llama 3.1 8B (Q8_0), expect roughly 24 tokens/sec. Speed scales with memory bandwidth (273 GB/s) more than anything else.
Can an NVIDIA DGX Spark (GB10) 128GB run a 70B model?
Yes — Llama 3.3 70B fits fully on-GPU (see the table for quant and speed).
Is the NVIDIA DGX Spark (GB10) 128GB good for local AI in 2026?
It runs 38 of the 51 models we track — a strong local-AI card. The full breakdown is in the table above.
Keep exploring
More GPU guides: RTX 5090 · RTX 5080 · RTX 5070 · RTX 5060 Ti 16GB · All GPU guides
Related reading: How to Run an LLM Locally (Beginner's Guide) · Q4 vs Q8 Quantization: VRAM Explained · Best Budget GPUs for Local AI