A small desktop box with 128GB of memory shared between the processor and the graphics chip, built to run large AI models at home. I run four of them joined together as one.
| Set-up | Model | Writing code | Writing prose | Reading (prefill) | Tested |
|---|---|---|---|---|---|
| Four DGX Sparks | GLM 5.3 FlashNVFP4 at 4 bits, dense layers at 8 | 104.1 tok/s | 62.7 tok/s | 3,159 tok/s | 11 Oct 2026 |
| Two DGX Sparks | Qwen 3.8 27BNVFP4 at 4 bits | 44.7 tok/s | 28.7 tok/s | 2,034 tok/s | 11 Oct 2026 |
| One DGX Spark | Qwen 3.8 27BNVFP4 at 4 bits | 26.7 tok/s | 16.9 tok/s | 1,188 tok/s | 11 Oct 2026 |
One user, the model's thinking switched off, the same prompts and the same script each time. My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W. All of my results, and how I test, are on the benchmarks page.
| Figure | What it is | From | Kind |
|---|---|---|---|
| $5,660 to $9,097 before tax | Other makers' 128GB versions of the DGX Spark, at a reseller in Germany10 October 2026, in DGX Spark vs RTX Spark laptop: ports, price, and what joins up | pi3g.com | store price |
| Figure | What it is | From | Kind |
|---|---|---|---|
| 58.4 to 94.2 tok/s | Tokens a second on code, before the recipe and after10 October 2026, in Four DGX Sparks twice as fast at ordinary writing: the recipe, and a real session | my own test | measured by me |
| 28.2 to 60.4 tok/s | Tokens a second on a short answer, before and after10 October 2026, in Four DGX Sparks twice as fast at ordinary writing: the recipe, and a real session | my own test | measured by me |
| 110.7 against 65.2 tok/s | Tokens a second in total with four requests at once, against the week before10 October 2026, in Four DGX Sparks twice as fast at ordinary writing: the recipe, and a real session | my own test | measured by me |
| 33.8 tok/s on prose, 58.4 on code | The same model writing on my four DGX Sparks, the middle of six runs3 October 2026, in M5 Ultra Mac Studio vs my four DGX Sparks: the same model | my own test | measured by me |
| 18 builds, 43,152 downloads | Builds I checked, and downloads in 30 days of the most downloaded one with the mismatch13 September 2026, in NVIDIA's NVFP4 builds of GLM 5.3 Flash: 18 checked for the fault | my own test | measured by me |
| about 20% | Speed lost with a build that also squeezes the activations to four bits13 September 2026, in NVIDIA's NVFP4 builds of GLM 5.3 Flash: 18 checked for the fault | my own test | measured by me |
| Fault gone, 99% of the speed | The weights-only build I moved to on 1 September13 September 2026, in NVIDIA's NVFP4 builds of GLM 5.3 Flash: 18 checked for the fault | my own test | measured by me |
| 42.8 tok/s | A 27 billion parameter model on two DGX Sparks6 September 2026, in RTX Spark: October launch, PAIR routing, a free RTX speed-up | my own test | measured by me |
| 46.3 tok/s | GLM 5.3 Flash on my four DGX Sparks, for a single user6 September 2026, in Qwen3.8 Flash Next and GLM 5.3 Flash on DGX Spark: first speeds | my own test | measured by me |
My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.
| Figure | What it is | From | Kind |
|---|---|---|---|
| 71 to 72 tok/s | Flash Next on one DGX Spark10 October 2026, in RTX Spark laptops are on sale: $2,599 to start, $5,899 for 128GB | forums.developer.nvidia.com | an owner's report |
| 90.3 tok/s on prose, 3,426 tok/s reading | The recipe author's own figures on his four DGX Sparks10 October 2026, in Four DGX Sparks twice as fast at ordinary writing: the recipe, and a real session | the recipe's page on GitHub | an owner's report |
| about £10,000 | Four DGX Sparks, going by what OpenFaaS paid28 September 2026, in Mac Studio M5 Ultra 512GB, and Apple's 3x cluster claim checked | blog.alexellis.io | an owner's report |
| 17.8 to 68.3 tok/s | The Pro version on eight Sparks, before and after speculative decoding28 September 2026, in MiMo-V2.6 tops the open-weights board, and runs on DGX Sparks | forums.developer.nvidia.com | an owner's report |
| about 45 tok/s for one user, 233 in total for 16 | The Flash version on two Sparks, for one user and for 16 users28 September 2026, in MiMo-V2.6 tops the open-weights board, and runs on DGX Sparks | forums.developer.nvidia.com | an owner's report |
| EXL3 at 3 bits a weight | The quant: 3 bits per weight, on the expert layers only19 September 2026, in DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock | forums.developer.nvidia.com | an owner's report |
| 25 to 40 tok/s for one user, 50 or more with several, up to 6 users | The two-Spark recipe: one stream, several streams, and the most users19 September 2026, in DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock | forums.developer.nvidia.com | an owner's report |
| 2GB (4GB if raised in the BIOS) | Memory a Spark reserves for the display19 September 2026, in DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock | forums.developer.nvidia.com | an owner's report |
| 1.75GB of the 2GB | How much of that the recipe gives the model19 September 2026, in DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock | forums.developer.nvidia.com | an owner's report |
| about £10,000, about £600 back | What OpenFaaS paid for four DGX Sparks, and what it got back19 September 2026, in Why a business bought 4 DGX Sparks: OpenFaaS's month with them | blog.alexellis.io | an owner's report |
The 10 most recent of 16; the rest are in the stories below. These are other people's figures on their own machines, with their own settings. I have not checked them and they should not be set side by side with mine.
| Figure | What it is | From | Kind |
|---|---|---|---|
| 695 tok/s | One DGX Spark answering 256 requests at once1 September 2026, in Two DGX Sparks or a Mac Studio Ultra: the same 256GB, the numbers side by side | see the story | a reviewer's test |
| Figure | What it is | From | Kind |
|---|---|---|---|
| 200Gb/s, up to four boxes | The DGX Spark's port for joining boxes10 October 2026, in DGX Spark vs RTX Spark laptop: ports, price, and what joins up | nvidia.com | the maker's claim |
| $4,999, from 23 October 2026 | The 64GB DGX Spark's on-sale date and starting price3 October 2026, in NVIDIA's 64GB DGX Spark at $4,999: what it runs | blogs.nvidia.com | NVIDIA's claim |
| 1.7 and 3.1 times | Two linked boxes and four, against one, in NVIDIA's table3 October 2026, in NVIDIA's 64GB DGX Spark at $4,999: what it runs | blogs.nvidia.com | NVIDIA's claim |
| 3 times | A cluster of Studios against a single Studio28 September 2026, in Mac Studio M5 Ultra 512GB, and Apple's 3x cluster claim checked | apple.com | Apple's claim |
| A 32K-token prompt on a 72B model, 4 machines against 1 | How Apple measured that 3x, from its footnote28 September 2026, in Mac Studio M5 Ultra 512GB, and Apple's 3x cluster claim checked | apple.com | the maker's claim |
| 552B | Parameters in DeepSeek V4.1 Flash19 September 2026, in DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock | huggingface.co | DeepSeek's model card |
| 552B and 196B | Parameters in the main network, plus the Engram lookup table13 September 2026, in DeepSeek V4.1 Flash on 4 DGX Sparks: the 552B open model | huggingface.co | DeepSeek's model card |
| October 2026 | When NVIDIA said RTX Spark PCs would ship6 September 2026, in RTX Spark: October launch, PAIR routing, a free RTX speed-up | blogs.nvidia.com | NVIDIA's statement |
| 20 cores, up to 1 petaflop, up to 128GB | The RTX Spark chip's CPU cores, AI compute and shared memory6 September 2026, in RTX Spark: October launch, PAIR routing, a free RTX speed-up | blogs.nvidia.com | NVIDIA's claim |
| 1.9 times | The most NVIDIA says local models speed up on RTX cards6 September 2026, in RTX Spark: October launch, PAIR routing, a free RTX speed-up | blogs.nvidia.com | NVIDIA's claim |
The 10 most recent of 14; the rest are in the stories below.
| Figure | What it is | From | Kind |
|---|---|---|---|
| $6,950 | NVIDIA's 128GB DGX Spark10 October 2026, in RTX Spark laptops are on sale: $2,599 to start, $5,899 for 128GB | theregister.com | reported |
| $7,000 to $9,000 | What a 128GB box costs, if you can find one3 October 2026, in NVIDIA's 64GB DGX Spark at $4,999: what it runs | tomshardware.com | reported |
| Figure | What it is | From | Kind |
|---|---|---|---|
| 2.7 tok/s | Llama 3.1 70B on one DGX Spark in its launch week1 September 2026, in Two DGX Sparks or a Mac Studio Ultra: the same 256GB, the numbers side by side | see the story | independent score |
| Figure | What it is | From | Kind |
|---|---|---|---|
| $1,000.99 | The desktop box over the 64GB DGX Spark10 October 2026, in RTX Spark laptops are on sale: $2,599 to start, $5,899 for 128GB | $5,999.99 minus $4,999 | my own maths |
| $1,050.01 | Microsoft's 128GB laptop under NVIDIA's own 128GB DGX Spark10 October 2026, in DGX Spark vs RTX Spark laptop: ports, price, and what joins up | $6,950 minus $5,899.99 | my own maths |
| $699.01 | Microsoft's 64GB laptop under NVIDIA's new 64GB DGX Spark10 October 2026, in DGX Spark vs RTX Spark laptop: ports, price, and what joins up | $4,999 minus $4,299.99 | my own maths |
| 2.50 times | Price per gigabyte, the 64GB version against the original at launch3 October 2026, in NVIDIA's 64GB DGX Spark at $4,999: what it runs | $78.11 divided by $31.24 | my own maths |
| 2,257 tok/s, 2.04 times | My four DGX Sparks reading a prompt, against the Mac Studio3 October 2026, in M5 Ultra Mac Studio vs my four DGX Sparks: the same model | 2,257 divided by 1,107 | my own maths |
| 0.89GB | Memory for a full 1,000,000 token window13 September 2026, in DeepSeek V4.1 Flash on 4 DGX Sparks: the 552B open model | 890 bytes times 1,000,000 tokens | my own maths |
Where mine came from. I bought my four DGX Sparks myself. NVIDIA has not supplied anything for this page and has not seen it.
Microsoft opened orders on Wednesday and five other makers follow on 16 October. The 128GB version is the one that runs the big models, and I would wait for independent tests before buying one.
I moved my four DGX Sparks to a recipe that another owner published on GitHub, and ordinary writing roughly doubled on the same hardware and the same model. The recipe is only a few weeks old, and these are one owner's figures from one cluster.
Microsoft's RTX Spark laptop now costs less than NVIDIA's own DGX Spark with the same memory. The DGX Spark is the one that can be joined to other boxes, and I do not know yet how many of the recipes its owners share will work on Windows.
NVIDIA announced a DGX Spark with 64GB of memory, half of what the original has, and it costs more than the full one did at launch. The price is the part that stood out to me, and I could not find how NVIDIA measured the gain it claims from linking two or four together.
MacStories tested GLM 5.3 Flash, the model my cluster runs, on a 256GB M5 Ultra Mac Studio, in a review I had missed. One Mac Studio writes about as fast as my four DGX Sparks and my Sparks read a prompt 2.04 times as fast, but the comparison is a rough one.
Apple says its latest Mac Studio can hold 512GB of memory in one box, the same as my whole cluster of four DGX Sparks, and that a cluster of Studios runs AI up to 3x faster than one. I looked up how Apple measured that, and I would wait for independent measurements before ordering.
Xiaomi's MiMo-V2.6-Pro sat at number one on Artificial Analysis's open-weights leaderboard on 27 September, and DGX Spark owners had both versions running within days. One owner reported 68.3 tokens a second on eight Sparks with speculative decoding, up from 17.8, and the early problem was tool use.
One of the regular recipe authors on NVIDIA's developer forum published a way to run DeepSeek V4.1 Flash on two DGX Sparks instead of four, with a second trick that applies to any headless Spark. I had not run it myself, and the speeds and the one quality check were the author's own.
OpenFaaS, a UK software company, has written up why it spent about £10,000 on four DGX Sparks: customer data it will not send to a cloud provider, and red teaming its own products. It also found two 4-bit builds of GLM 5.3 Flash running at different speeds, so two builds of the same model are not interchangeable.
DeepSeek released a 552 billion parameter model with open weights on 10 September, and within a day an owner had it running on four DGX Sparks. The speeds were similar to GLM 5.3 Flash on my own four, but they came from one owner's first day on a four bit build that had not been checked for quality, as far as I could see.
NVIDIA published its own NVFP4 build of GLM 5.3 Flash on 9 September. I checked 18 builds of that model and of Qwen 3.8 Flash Next for the fault I had found on my own cluster, and found it in three GLM builds, the most downloaded one among them, and in neither of NVIDIA's own.
NVIDIA confirmed at IFA in Berlin that RTX Spark PCs would ship in October, and released PAIR, a free tool that spreads AI work across the PCs on a home network. Nobody outside NVIDIA had benchmarked an RTX Spark, and I think the free speed-up for RTX cards mattered more on the day.
By 6 September owners on NVIDIA's forum had Qwen3.8 Flash Next running on a single DGX Spark at 37 to 44 tokens a second, and one post had GLM 5.3 Flash on one box at 64. The 64 came from a build squeezed to about 2 bits per weight, and I would want to see the quality numbers before copying it.
A buyer with two new DGX Sparks asked on Reddit whether to send them back now that Apple had announced the M5 Ultra Mac Studio. I put my own cluster's numbers next to what Mac Studio owners measure, and I think the answer depends on the work: short questions suit the Mac, and agents and long documents suit the Sparks.