NVIDIA DGX Spark for Local AI (2026): Price, 64 GB vs 128 GB, Real Speeds and Setup

The NVIDIA DGX Spark is a 1.2 kg box that puts 128 GB of unified memory and a Blackwell GPU on your desk. It can load models no single gaming GPU can hold, like gpt-oss-120b, and it runs the full CUDA stack.

It is also slower at generating text than you might expect, and its price has gone from $3,999 at launch to $6,950 for the 128 GB model in October 2026.

This guide covers the specs, the new 64 GB model, real benchmark numbers, what fits in memory, how to set it up, and who should buy one instead of a GPU or a Mac.

DGX Spark Specs

These are from the official DGX Spark datasheet. Every DGX Spark and partner GB10 box shares the same chip; partners mainly change storage, case and price.

DGX Spark Price in 2026 (and Why It Went Up)

NVIDIA blamed both rises on the global memory shortage. AI data centres are buying so much DRAM that LPDDR5x and DDR5 prices have jumped, and Micron has warned supply will stay tight into 2028. My guide to why RAM is so expensive in 2026 explains the shortage in detail.

Partner boxes using the same GB10 chip come from Acer, ASUS, Dell, Gigabyte, HP and MSI. In February, Tom’s Hardware found the ASUS Ascent GX10 (1 TB SSD instead of 4 TB) for about $3,267. Partner prices have moved with memory costs since then, so check today’s price before you buy.

Tip: the smaller SSD is the main difference on most partner boxes. Model files are big, but a 2 TB external USB-C SSD costs far less than the price gap, so a cheaper partner unit is often the better deal.

64 GB vs 128 GB: Which Model?

The 64 GB version launched on 23 October 2026 through partners only; there is no NVIDIA-branded 64 GB unit. It keeps the same GB10 chip, the same 273 GB/s bandwidth and the ConnectX-7 port, but halves the memory and the onboard storage.

The whole point of a Spark is memory. If 64 GB is enough for your models, a 32 GB GPU or a 64 GB Mac usually gives you faster tokens for less money. I’d only buy the Spark in its 128 GB form.

Real Benchmarks: Tokens per Second

I haven’t tested a Spark myself, so these numbers come from the llama.cpp project’s own DGX Spark benchmark file (llama-bench, build 7941, February 2026) and the LMSYS review. Prompt processing (pp) is how fast it reads your prompt; text generation (tg) is how fast it writes the answer.

Two patterns stand out. MoE models (gpt-oss, Qwen3-Coder 30B-A3B, GLM-4.7-Flash) run well, because only a few billion parameters are read per token. Dense models are slow: a 7B dense model at Q8 generates only about 29 t/s, and LMSYS measured Llama 3.1 70B FP8 at just 2.7 t/s.

For comparison, LMSYS measured gpt-oss-20b in Ollama at 49.7 t/s on the Spark, 205 t/s on an RTX 5090 and 215 t/s on an RTX PRO 6000 Blackwell. Software has improved since then (llama.cpp now reaches 83 t/s on the same model), but discrete GPUs remain far ahead when the model fits in their VRAM.

Note: prompt processing is the Spark’s strength. Reading a long document or a big code base at 2,000-4,500 tokens per second makes agents and RAG feel much quicker than on a Mac with similar memory.

Which Models Fit in 128 GB and 64 GB

Leave about 8-12 GB for the OS, the inference server and context. Sizes are typical GGUF file sizes.

For VRAM maths on any model, use my VRAM requirements guide. If you want the very large models, my RAM for local AI guide shows how far system RAM offload can take a normal PC.

DGX Spark vs RTX 5090 vs Mac Studio vs Ryzen AI Max

Tom’s Hardware found the Spark’s GPU has roughly twice the horsepower of the Ryzen AI Max+ 395’s Radeon 8060S, and much better long-context and image results. The Ryzen boxes cost less though.

My Mac vs PC for local AI guide goes deeper on the Mac and Ryzen side, and RTX 5090 vs RTX 5080 for AI covers the GPU route.

How to Set Up a DGX Spark for Local AI

The Spark ships with DGX OS, NVIDIA’s Ubuntu with drivers, CUDA and Docker already installed. You can use it with a monitor and keyboard or headless from another computer.

1. First boot: connect HDMI, keyboard and mouse, or follow the headless setup in NVIDIA’s quick start guide, then create your user.
2. Update the system: run sudo apt update && sudo apt full-upgrade and reboot. NVIDIA ships frequent performance updates.
3. Connect from your PC or Mac: install NVIDIA Sync, or simply SSH in with ssh user@spark-ip.
4. Install Ollama with curl -fsSL https://ollama.com/install.sh | sh, then try ollama run gpt-oss:120b.
5. Or build llama.cpp with CUDA for the best speed: cmake -B build -DGGML_CUDA=ON then cmake --build build -j.
6. Add a chat UI: run Open WebUI in Docker and point it at Ollama. My Open WebUI setup guide has the commands.
7. Use the NVIDIA playbooks at build.nvidia.com/spark for ready-made setups (vLLM, SGLang, ComfyUI, fine-tuning).

Warning: the CPU is Arm, not x86. Most AI tools have Arm64 Linux builds, but some Python wheels and Docker images don’t. Look for arm64 or aarch64 images first.

Not sure whether to use Ollama, vLLM or llama.cpp? My vLLM vs llama.cpp vs Ollama comparison explains which one fits which job. On the Spark, vLLM or SGLang make sense when several users or agents share it, because batching uses the GPU far better.

Image, Video and Fine-Tuning

• ComfyUI: works on the Spark because it has CUDA, and the 128 GB lets you load big image and video models without offloading. It is still slower than an RTX 5090 per image, since diffusion also depends on raw GPU speed.
• Fine-tuning: this is a strong use. 128 GB is enough for LoRA and QLoRA training on models that would never fit on a 24-32 GB card.
• FP4: Blackwell supports NVFP4, which shrinks models and speeds them up. My NVFP4 vs MXFP4 guide explains the formats.

Linking Two Sparks

The ConnectX-7 ports let you join two Sparks with one QSFP cable. NVIDIA says a pair can run models of up to about 405B parameters in FP4.

Two 64 GB units cost about $10,000 for a 128 GB pool, which is more than one 128 GB unit at $6,950. Linking only makes sense with two 128 GB machines, and even then the network link adds overhead, so expect lower speed per token than on one box.

Who Should Buy One

Troubleshooting

Ollama is much slower than the benchmarks

Update DGX OS and Ollama, then check the model is on the GPU with ollama ps (it should say 100% GPU). Many early slow reports were fixed by updates. llama.cpp built with CUDA is often faster than Ollama on the same model.

A model doesn’t load even though it should fit

The 128 GB is shared with the OS and CPU apps. Close other programs, reduce the context length, or use a smaller quant. Check free memory with free -h.

A Python package or Docker image fails to install

It probably has no Arm64 build. Look for an aarch64 image, use NVIDIA’s NGC containers, or build from source.

Dense 70B models are painfully slow

That is the bandwidth limit, not a fault. Switch to an MoE model such as gpt-oss-120b, which is both smarter and many times faster on the Spark.

The box gets hot or loud

Give it space on all sides and don’t stack anything on top. Reviewers found it handles sustained load well when it has airflow.

Frequently Asked Questions

How much does the DGX Spark cost?

The 128 GB model is $6,950 in the US as of October 2026, up from $3,999 at launch. New 64 GB partner models start at $4,999.

Is the DGX Spark good for running LLMs locally?

Yes for large MoE models such as gpt-oss-120b, which runs at about 59 tokens per second in llama.cpp. It is slow on big dense models, and a fast GPU beats it on anything that fits in 32 GB of VRAM.

What is the DGX Spark’s memory bandwidth?

273 GB/s. That is far below an RTX 5090 at about 1.8 TB/s and is the main reason its text generation is slower.

Can the DGX Spark run ComfyUI?

Yes. It has full CUDA support and enough memory for large image and video models, but it renders slower than an RTX 5090.

Is the DGX Spark better than a Mac Studio?

It is better if you need CUDA, fast prompt processing or fine-tuning. A Mac Studio has more memory bandwidth and can hold larger models, so it often generates text faster.

Can you game on a DGX Spark?

Not really. It runs DGX OS on an Arm CPU and isn’t designed for games.

Should I buy the 64 GB DGX Spark?

Usually not. At $4,999 it costs more than a 32 GB GPU PC that runs the same-size models faster. The 128 GB version is where the Spark makes sense.

Final Thoughts

The DGX Spark is a CUDA development box with a lot of memory, not a speed monster. If you build with 100B+ MoE models, fine-tune, or need the same software stack as NVIDIA’s data centre GPUs, it does a job no gaming card can.

If you mainly want fast local chat or fast image generation, the price rise has made it a hard sell. A 24-32 GB GPU is quicker for those jobs, and my best GPU for local LLMs guide lists the current picks by budget.

Sources

• llama.cpp: DGX Spark benchmark results
• LMSYS: NVIDIA DGX Spark in-depth review
• PNY: NVIDIA DGX Spark datasheet
• Tom’s Hardware: DGX Spark price rises to $4,699 (February 2026)
• Hardware Busters: DGX Spark 64 GB at $4,999, 128 GB at $6,950 (October 2026)
• Tom’s Hardware: DGX Spark review
• NVIDIA Developer Forums: price change announcement