Best AI Models for NVIDIA DGX Spark
128 GB total — ~120 GB usable as VRAM
128 GB coherent unified memory at a modest 273 GB/s — huge capacity but bandwidth-bound, so very large models load yet generate at low tokens/sec.
With 128 GB of memory, this is a high-end configuration for local AI. You can comfortably run most open-source LLMs including large 70B parameter models at good quantization levels, making it one of the best setups for serious local AI work.
At this memory tier, nearly every popular open-source model is within reach. You can run Llama 3 70B at Q4_K_M or even Q5_K_M quantization with room to spare, handle coding assistants like DeepSeek Coder 33B at high quality, and easily run any 7B–30B model at full or near-full precision. Context windows remain generous even with larger models, so multi-turn conversations and long-document processing work smoothly.
Runs Well
- 70B models (Llama 3 70B, Qwen 72B) at Q4–Q5
- 30B models at Q6–Q8 quality
- 7B–14B models at full FP16 precision
- Vision models (LLaVA, CogVLM) without compromise
Challenging
- Mixture-of-experts models like Mixtral 8x22B at higher quants
- 120B+ models still require lower quantizations
What LLMs Can NVIDIA DGX Spark Run?
152 models · 15 excellent · 39 good
Showing compatibility for NVIDIA DGX Spark
| Model | Quant | VRAM | Speed | Context | Status | Grade |
|---|---|---|---|---|---|---|
Q4_K_M·2.1 t/s tok/s·66K ctx·BARELY RUNS | Q4_K_M | 85.1 GB | 2.1 t/s | 66K | BARELY RUNS | C38 |
Q4_K_M·2.1 t/s tok/s·66K ctx·BARELY RUNS | Q4_K_M | 85.1 GB | 2.1 t/s | 66K | BARELY RUNS | C38 |
Q4_K_M·2.4 t/s tok/s·131K ctx·BARELY RUNS | Q4_K_M | 72.7 GB | 2.4 t/s | 131K | BARELY RUNS | C38 |
Q4_K_M·2.4 t/s tok/s·262K ctx·BARELY RUNS | Q4_K_M | 74.7 GB | 2.4 t/s | 262K | BARELY RUNS | C38 |
Q4_K_M·1.8 t/s tok/s·1049K ctx·BARELY RUNS | Q4_K_M | 99.5 GB | 1.8 t/s | 1049K | BARELY RUNS | C34 |
Q4_K_M·2.5 t/s tok/s·BARELY RUNS | Q4_K_M | 71.7 GB | 2.5 t/s | — | BARELY RUNS | C38 |
IQ2_XXS·1.8 t/s tok/s·131K ctx·BARELY RUNS | IQ2_XXS | 99.2 GB | 1.8 t/s | 131K | BARELY RUNS | C34 |
Q4_K_M·2.3 t/s tok/s·BARELY RUNS | Q4_K_M | 78.5 GB | 2.3 t/s | — | BARELY RUNS | C38 |
Q4_K_M·2.7 t/s tok/s·131K ctx·BARELY RUNS | Q4_K_M | 66.7 GB | 2.7 t/s | 131K | BARELY RUNS | C38 |
BF16·2.4 t/s tok/s·262K ctx·BARELY RUNS | BF16 | 72.6 GB | 2.4 t/s | 262K | BARELY RUNS | C38 |
Q4_K_M·2.4 t/s tok/s·BARELY RUNS | Q4_K_M | 73.3 GB | 2.4 t/s | — | BARELY RUNS | C38 |
Q4_K_M·2.7 t/s tok/s·131K ctx·BARELY RUNS | Q4_K_M | 65.1 GB | 2.7 t/s | 131K | BARELY RUNS | C38 |
Q4_K_M·216.4 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 0.8 GB | 216.4 t/s | 131K | RUNS GREAT | S100 |
Q4_K_M·3.8 t/s tok/s·131K ctx·BARELY RUNS | Q4_K_M | 46.6 GB | 3.8 t/s | 131K | BARELY RUNS | C40 |
Q4_K_M·3.6 t/s tok/s·262K ctx·BARELY RUNS | Q4_K_M | 49.2 GB | 3.6 t/s | 262K | BARELY RUNS | C40 |
Q4_K_M·51.7 t/s tok/s·131K ctx·RUNS WELL | Q4_K_M | 3.4 GB | 51.7 t/s | 131K | RUNS WELL | A80 |
NVIDIA DGX Spark Specifications
- Brand
- NVIDIA
- Chip
- GB10
- Type
- AI Box
- Unified Memory
- 128 GB
- Memory Bandwidth
- 273.0 GB/s
- GPU Cores
- 6144
- CPU Cores
- 20
- Neural Engine
- 1000.0 TOPS
- Form Factor
- mini (150x150x50.5mm, 1.2kg)
- Architecture
- Blackwell
- Memory Type
- LPDDR5X
- TDP
- 140 W
- PSU
- 240 W
- MSRP
- $3,999
- Release Date
- 2025-10-15
Get Started
Prompt Processing
Estimated for SOLAR 10.7B Instruct v1.0, the compute-bound phase that reads your prompt before the first reply token appears.
Short chat
195 ms
512 tok prompt
Long chat
1.6 s
4,096 tok prompt
Document / codebase
12.5 s
32,768 tok prompt
Prefill is compute-bound and a different number from the decode tok/s shown elsewhere on this page — how prompt processing works →
Performance figures are estimates calibrated as of 2026-07-30 — see calibration basis →
Devices to Consider
Similar devices and upgrades with more memory or higher bandwidth
Frequently Asked Questions
- Can NVIDIA DGX Spark run Mixtral 8x22B v0.1?
Yes, the NVIDIA DGX Spark with 128 GB unified memory can run Mixtral 8x22B v0.1, WizardLM 2 8x22B, GPT OSS 120B, and 1720 other models. 438 models achieve excellent performance, and 523 run at good quality. Apple Silicon's unified memory architecture lets the GPU access the full memory pool without copying data, making it efficient for AI workloads.
- How much memory is available for AI on NVIDIA DGX Spark?
The NVIDIA DGX Spark has 128 GB unified memory. After macOS reserves ~3.5 GB for the operating system, approximately 124.5 GB is available for AI models. Unlike discrete GPUs where VRAM is separate from system RAM, Apple Silicon shares one memory pool between the CPU and GPU — this means no data copying overhead, but you share memory with macOS and open apps.
- Is NVIDIA DGX Spark good for AI?
With 128 GB unified memory and 273.0 GB/s bandwidth, the NVIDIA DGX Spark is excellent for running local AI models. It supports 961 models at good quality or better. This is a premium configuration — you can run large 30B+ parameter models at good quality, and most 7B models at maximum quality. Ideal for professional AI workloads.
- What's the best model for NVIDIA DGX Spark?
The top-rated models for the NVIDIA DGX Spark are Mixtral 8x22B v0.1, WizardLM 2 8x22B, GPT OSS 120B. With this much memory, you can prioritize quality — use higher quantizations (Q5/Q6) for better output, or run larger 30B+ models for more capable reasoning.
- How fast is NVIDIA DGX Spark for AI inference?
With 273.0 GB/s memory bandwidth, the NVIDIA DGX Spark achieves approximately 43 tok/s on a 7B model at Q4_K_M — that's very fast, well above conversational speed. A 14B model runs at ~21 tok/s. Apple Silicon achieves high efficiency (~70%) thanks to unified memory — there's no PCIe bottleneck between CPU and GPU.
tok/s = (273 GB/s ÷ model GB) × efficiency
Apple Silicon achieves ~70% bandwidth efficiency thanks to unified memory and Metal acceleration.
Estimated speed on NVIDIA DGX Spark
~2 tok/s~2 tok/s~2 tok/s~2 tok/sReal-world results typically within ±20%.
- Can I run AI offline on NVIDIA DGX Spark?
Yes — once you download a model, it runs entirely on the NVIDIA DGX Spark without internet. Applications like Ollama and LM Studio make it straightforward to download, manage, and run models locally. All your conversations stay private on your device with zero data sent to external servers. This is one of the key advantages of local AI: complete privacy, no API costs, and no rate limits.
- Anything to watch out for with NVIDIA DGX Spark?
128 GB coherent unified memory at a modest 273 GB/s — huge capacity but bandwidth-bound, so very large models load yet generate at low tokens/sec.