NVIDIA DGX Spark for local AI: CUDA, 128 GB and very fast prompts (2026)

By /
NVIDIA DGX Spark page on NVIDIA UK, October 2026

The NVIDIA DGX Spark is a small Linux desktop built on the GB10 chip, with 128 GB of memory shared by the CPU and GPU. NVIDIA sells one size only, so there is no memory choice to make. It runs Llama 3.3 70B and gpt-oss-120b, but big dense models write slowly. Expect to pay £5,249 at Scan. It is the simplest way to get CUDA with lots of memory. Our rating is ★★★½☆ 3.5/5. Still choosing? See which Mac or mini PC to buy for local AI.

Last updated 22 October 2026. Written by Annie, a search and digital marketing specialist. Based on the maker's specifications, UK prices and published benchmarks; I have not tested this machine myself. See the editorial policy.

NVIDIA DGX Spark at a glance for local AI: £5,249 with 128 GB and 4 TB, runs 70B models with very fast prompt processing, rated 3.5 out of 5
NVIDIA DGX Spark at a glance. Prices from UK stores, checked 17 October 2026.

Key facts

  • What it is: NVIDIA's own GB10 desktop (the Founders Edition), running NVIDIA DGX OS, an Ubuntu-based Linux.
  • Chip: NVIDIA GB10 Grace Blackwell: a 20-core Arm CPU (10 Cortex-X925 and 10 Cortex-A725) and a Blackwell GPU rated at up to 1 PFLOP at FP4.
  • Memory options: 128 GB LPDDR5X unified memory only. It is fixed at purchase; there is no memory upgrade. A 64 GB version is coming, but only from partner brands.
  • Usable for AI: there is no fixed GPU share. The GPU can use almost all of the 128 GB. In one owner's tests the OS and desktop took about 6-7 GB, so plan on roughly 120 GB for models.
  • Memory bandwidth: 273 GB/s
  • UK price: from £5,249 (128 GB, 4 TB SSD) at Scan, as checked 17 October 2026. NVIDIA's own UK store charges £6,390.
  • Where to buy: Scan (in stock), NVIDIA Marketplace UK, plus business resellers on NVIDIA's partner list.
  • Best for: developers who need CUDA and NVIDIA's software, and anyone feeding long documents or code into mixture-of-experts models.
  • Our rating (desk review, not hands-on): ★★★½☆ 3.5/5

On this page

Which AI models can the DGX Spark run?

With roughly 120 GB free for models, the Spark holds anything up to about 190B parameters at 4-bit (NVIDIA says up to 200B).

Memory optionBiggest model at 4-bitExample modelsPublished speed
128 GB (the only NVIDIA version)About 190B dense; Llama 3.3 70B even fits at 8-bit (74.2 GB)Llama 3.3 70B (42.4 GB), gpt-oss-120b (59 GiB), Qwen3 32B (19.4 GB)Llama 3.1 70B, Q4_K_M, Ollama: 4.4 tokens a second (Ollama). gpt-oss-120b, MXFP4, llama.cpp: 58.7 tokens a second (llama.cpp)
64 GB (partner machines only, on sale from 23 October 2026)About 56 GB is left for models, so about 85B; Llama 3.3 70B fits with a short conversationQwen3.8 27B (16.4 GB), Qwen3 32B (19.4 GB)No published figure

Long conversations slow it down: in llama.cpp, gpt-oss-120b falls from 58.7 to 42.8 tokens a second once 32,000 tokens of context are loaded. Qwen3 235B-A22B (142.5 GB at 4-bit) does not fit in one Spark, but two linked Sparks ran Qwen3-VL-235B-A22B at about 13 tokens a second in a llama.cpp test.

How fast is it?

The Spark has a split personality. It reads prompts very fast, but it writes answers at the pace its 273 GB/s memory allows. LMSYS, which tested it at launch, named memory bandwidth as the main bottleneck.

Published figures, prompt processing first, then generation:

  • gpt-oss-120b, MXFP4, llama.cpp (official llama.cpp benchmark, February 2026, 2,048-token prompt): 2,444 tokens a second reading, 58.7 writing (llama.cpp).
  • Qwen3 Coder 30B-A3B, Q8_0, llama.cpp (same run): 2,987 reading, 61.1 writing.
  • Qwen2.5 Coder 7B, Q8_0, llama.cpp (same run): 2,250 reading, 29.4 writing.
  • Llama 3.1 70B, Q4_K_M, Ollama 0.12.6 (October 2025): 1,911 reading, 4.4 writing (Ollama).
  • Qwen3 32B, Q4_K_M, Ollama 0.12.6: 705 reading, 9.4 writing.
  • Llama 3.1 70B, FP8, SGLang (LMSYS, October 2025): 803 reading, 2.7 writing (LMSYS).
  • GPT-OSS 20B, MXFP4, Ollama, batch 1 (LMSYS): 2,053 reading, 49.7 writing.

That gap is the thing to understand before you buy. For Llama 3.1 70B, Ollama measured prompt reading about 430 times faster than writing. A long document goes in quickly, then the answer arrives at a few words a second. Mixture-of-experts models such as gpt-oss-120b avoid most of the pain, because they only read a fraction of their weights for each token.

Compared with AMD's Ryzen AI Max+ 395, as used in the HP Z2 Mini G1a and Framework Desktop, writing speed is about the same. kyuz0's llama.cpp tests on a Framework Desktop (May 2026, Vulkan) got 56.6 tokens a second writing with gpt-oss-120b, against the Spark's 58.7. Prompt reading is where the Spark wins: about 720 tokens a second on the AMD chip (512-token prompt) against 2,444 on the Spark (2,048-token prompt). The tests differ slightly, so treat that as roughly three times faster, not an exact figure.

Which version should you buy?

There is only one NVIDIA version: 128 GB with a 4 TB SSD. Buy that one if you want the Spark itself.

If you don't need 4 TB of storage, look at the ASUS Ascent GX10 first. It is the same chip and memory with a 1 TB drive, from £4,680. Skip the 64 GB partner models unless your models are 32B or smaller. NVIDIA's US price for the 64 GB version is from $4,999 (about £3,760, not a UK price), which is poor value next to 128 GB.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
DGX Spark128 GB4 TB£5,249.00Scan
DGX Spark, open box128 GB4 TB£5,159.99Scan
DGX Spark128 GB4 TB£6,390.00NVIDIA Marketplace UK
DGX Spark with 3-year NVIDIA Enterprise128 GB4 TB£7,178.00Scan
Two Sparks plus a 200G cable2 x 128 GB2 x 4 TB£10,649.00Scan

All prices include VAT and were checked on 17 October 2026. Scan's price is well below NVIDIA's own UK price, so it may not last. Scan also rents remote access to a Spark from £69.98 a week through Scan Cloud, a cheap way to try your models first.

Hidden costs

  • Memory you can't add later. 128 GB is all you will ever have. The upside is that the 4 TB SSD is included, so you won't need to pay for storage.
  • Rising prices. NVIDIA raised the US price from $3,999 to $4,699 in February 2026, blaming memory supply. It now lists $6,950 in the US and £6,390 in the UK. Scan sold it for £3,699.98 in February.
  • Short warranty. NVIDIA gives one year. There is no extended NVIDIA cover for home buyers on the pages I checked.
  • Power. After NVIDIA's February 2026 update it idles at about 25 W. Left on all year, that is about £58 at the October 2026 price cap rate of 26.32p per kWh.
  • A screen and keyboard, unless you run it headless and connect from another computer, as LM Studio's guide describes.

Software: Ollama, LM Studio and the rest

This is the Spark's strongest point. It ships with DGX OS and NVIDIA's AI software preinstalled, and CUDA 13 runs on it (llama.cpp's February 2026 test shows CUDA 13.0 with driver 580.95.05).

  • Ollama lists the GB10 as a supported GPU, and NVIDIA has a step-by-step guide for Ollama with Open WebUI.
  • LM Studio was a launch partner. It released a Linux Arm build with a CUDA 13 version of its llama.cpp engine for the Spark. You can use it on the Spark or serve models to other computers on your network.
  • llama.cpp, vLLM, SGLang and TensorRT-LLM all have NVIDIA playbooks, as do ComfyUI and fine-tuning with Unsloth, LLaMA Factory and NeMo.
  • What doesn't run: Windows and Mac-only tools such as MLX. It is an Arm Linux machine.
  • Setup steps: install system updates through the DGX Dashboard to get the idle power fix. Don't trust nvidia-smi's memory figure; NVIDIA says it shows "Not Supported" on this chip, and some memory checks under-report what you can actually use.

Noise, power and size

  • Size: 150 x 150 x 50.5 mm and 1.2 kg. It is powered by a 240 W external USB-C supply; the GB10 chip itself is rated at 140 W.
  • Noise: NVIDIA declares 29 dB(A) at the operator position under maximum GPU load and 13 dB(A) at idle. ServeTheHome found it hard to hear from 1-1.5 m, and it stayed under 40 dBA outside stress tests.
  • Measured power: ServeTheHome saw 40-45 W idle at launch, 60-90 W in many AI inference jobs and just under 200 W under full stress. Tom's Hardware measured idle falling from about 37 W to 25 W after NVIDIA's February 2026 update (22 W with no display attached).
  • Heat: in StorageReview's test serving gpt-oss-120b with vLLM, the Founders Edition GPU peaked at 80-82°C, in line with the Dell, Gigabyte and ASUS versions.

Who makes the DGX Spark?

  • NVIDIA, founded on 5 April 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem, and based in Santa Clara, California.
  • The GB10 chip was developed with MediaTek.
  • Support in the UK is through NVIDIA Customer Care online (knowledge base, questions and live chat). There is no UK walk-in service.
  • Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI sell their own GB10 machines alongside NVIDIA's.

What's good

  • CUDA and NVIDIA's full AI stack work out of the box, with step-by-step playbooks.
  • Prompt reading is about three times faster than AMD's Ryzen AI Max+ 395 on gpt-oss-120b.
  • 128 GB runs gpt-oss-120b at about 59 tokens a second in llama.cpp.
  • Small and quiet: 1.2 kg, with a declared 29 dB(A) at full GPU load.
  • A 200 Gb/s ConnectX-7 port lets you link two Sparks for models too big for one.
  • A 4 TB SSD is included.

What's not so good

  • Big dense models write slowly: 4.4 tokens a second for Llama 3.1 70B at 4-bit in Ollama.
  • The price has climbed sharply in 2026, from £3,699.98 at Scan in February to £5,249 today, and £6,390 direct from NVIDIA.
  • Linux on Arm only. No Windows.
  • One-year warranty.
  • Idle power was high at launch. The fix only arrived in February 2026.

Who should buy it?

Buy the Spark if you build AI software and need CUDA, or if your work means pasting long documents or codebases into a model. It is also a good fit for mixture-of-experts models. If you mostly want to chat with a 70B model, the slow writing speed will frustrate you, and a machine with more memory bandwidth will feel quicker. If you want the same chip for less, buy the ASUS Ascent GX10.

Buying in the UK

  • Availability: Scan, based in Bolton, has it in stock with free delivery, a next-day option and free collection from its Bolton store. NVIDIA's UK store sells it too, with a free NVIDIA training course.
  • VAT: UK prices include VAT. Importing one from the US at $6,950 would cost about £5,226 before 20% import VAT and any duty, which GOV.UK says the courier collects on goods over £135. That makes it dearer than Scan.
  • Warranty: NVIDIA gives one year on new units (three months on refurbished ones). It covers only units bought from NVIDIA or an authorised reseller, and it can't be transferred. Scan lists the warranty as "Direct", so claims go to NVIDIA.
  • Returns: Scan's terms give you 14 days from delivery to cancel, and you pay the return postage. Goods faulty on arrival can be refunded if returned within 28 days.

Interesting facts about the DGX Spark

  • Using a Spark in a large data centre or GPU cluster voids NVIDIA's warranty.
  • Tom's Hardware reports a Spark deployment for the IceCube project at the South Pole, where its low power draw matters.
  • Scan rents remote access to Sparks by the week, month or year, so you can test your models before spending £5,000.

Alternatives to the DGX Spark

  • ASUS Ascent GX10 - the same GB10 chip and 128 GB with a 1 TB SSD, from £4,680. Pick it if you don't need 4 TB.
  • HP Z2 Mini G1a - AMD's Ryzen AI Max+ PRO 395 with up to 128 GB, Windows and a UK on-site warranty. Pick it if you want Windows and don't need CUDA.
  • Framework Desktop - the same AMD chip as the HP in a repairable desktop. Pick it if you want similar writing speed for chat and don't need CUDA.

Our rating

AreaScore
Memory for AI4/5
Speed3/5
Value2/5
Software5/5
UK buying4/5
Overall★★★½☆ 3.5/5

The best software of any local AI box, in a tiny, quiet case. It is let down by slow writing on big dense models and a price that has risen by about £1,500 at Scan this year.

Frequently asked questions

Can the DGX Spark run Llama 70B?

Yes. Llama 3.3 70B at 4-bit needs about 42.4 GB, so it fits easily. Ollama measured Llama 3.1 70B at 4-bit writing 4.4 tokens a second, which is slow for chat.

How much does the DGX Spark cost in the UK?

£5,249 at Scan, checked 17 October 2026. NVIDIA's own UK store charges £6,390.

Is the DGX Spark better than the ASUS Ascent GX10?

They use the same chip and memory, and reviewers found only small speed differences between GB10 machines. The Spark has a 4 TB SSD and NVIDIA's own warranty; the GX10 is cheaper with 1 TB.

Is the DGX Spark better than a Mac mini or Mac Studio for local AI?

It depends on what you run. The DGX Spark has 128 GB of unified memory at 273 GB/s and NVIDIA's CUDA software, from £5,249 in the UK, and two Sparks can be linked for models too big for one. A Mac mini M5 Pro tops out at 64 GB, with about 48 GB available to models by default, so it suits small and mid-sized models for far less money. A Mac Studio M5 Ultra offers 96 GB, 256 GB or 512 GB, so it runs bigger models than one Spark, but without CUDA. Pick the Spark if your tools need NVIDIA; pick a Mac if you want macOS and the most memory.

Sources

More guides

Scroll to Top