DGX Spark/Guides/DGX Spark 64GB vs 128GB
Buying guide · November 2026

DGX Spark 64GB vs 128GB: same speed, half the memory, and which models still fit

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

The 64 GB DGX Spark uses the same GB10 chip and the same 273 GB/s of memory bandwidth as the 128 GB model, so a model that fits runs at the same speed; it only holds half as much. NVIDIA announced it on 2 October 2026 from $4,999, on sale from 23 October through Acer, ASUS, Dell, Gigabyte, HP and MSI only, against $6,950 for the 128 GB Founders Edition. By my calculation a 64 GB unit leaves about 54 GiB for weights and KV cache: enough for 20B to 35B models at long context, or a 4-bit 70B up to about 32,000 tokens, but not for gpt-oss-120b. Prices were checked on 7 November 2026.

Item64 GB128 GBSource
Chip and computeGB10, up to 1 PFLOP FP4 (sparse)SameNVIDIA spec table
Memory bandwidth273 GB/s, 256-bit273 GB/s, 256-bitNVIDIA, The Register
NVIDIA's model-size guidance, one unitUp to 100B parametersUp to 200B parametersNVIDIA
Two linked unitsUp to 200B parametersUp to 400B parametersNVIDIA
Start price (US)From $4,999$6,950 (Founders Edition)NVIDIA blog; price history
Who sells itAcer, ASUS, Dell, Gigabyte, HP, MSINVIDIA and all GB10 partnersNVIDIA blog
Left for weights and KV cacheAbout 53.6 GiB (my estimate)About 117.6 GiB (my estimate)Worked below

Spec rows read on NVIDIA's product page on 7 November 2026. The page lists one bandwidth and one compute figure for both memory sizes. The last row is my arithmetic, shown in the memory section.

Is the 64 GB DGX Spark slower than the 128 GB one?

No. Every first-party figure I could find gives the 64 GB model the same memory bandwidth. NVIDIA's spec table lists "64 GB LPDDR5X or 128 GB LPDDR5x" against a single 256-bit interface and 273 GB/s. The Register reported on 2 October 2026 that bandwidth is unchanged, which points to lower-capacity LPDDR5x packages rather than fewer of them. The ASUS announcement summarized by igor'sLAB gives the 64 GB Ascent GX10 273 GB/s, and Dell's configurator lists both of its options at "up to 8533 MT/s".

That matters because decode speed on this chip is set by bandwidth: tokens per second is at most 273 GB/s divided by the bytes of weights read per token. On a 128 GB node, vLLM 0.27.1 decoded Llama 3.3 70B AWQ at 6.0 tok/s for one user in GPUwerk's own run of 2 September 2026 (benchmarks). A 64 GB unit reads the same 37 GiB per token from the same memory bus, so I expect the same figure when the model fits. That is an inference: nobody I could find has published a 64 GB benchmark yet. Prefill uses the same GB10 compute, so it should not differ either.

Who sells the 64 GB model, and what does it cost?

Six partners, no NVIDIA store. NVIDIA's blog names Acer, ASUS, Dell, Gigabyte, HP and MSI and a start price of $4,999, and its product page says the 64 GB configuration is "available exclusively through participating OEM partners". Lenovo is not on the list, so the ThinkStation PGX stays 128 GB only. These are the listings I could read on 7 November 2026:

MachineConfigurationPriceSourceChecked
Any 64 GB GB10 (NVIDIA guidance)64 GBFrom $4,999NVIDIA blog2026-11-07
Dell Pro Precision with GB10 (FCM1253), US128 GB, 2 TB$7,901dell.com US2026-11-07
Same, 64 GB option64 GB, 2 TB$5,751 (64 GB option is minus $2,150)dell.com US2026-11-07
Same, 64 GB with 1 TB64 GB, 1 TB$5,371 (my sum of the listed deltas)dell.com US2026-11-07
Dell Pro Precision with GB10, Ireland128 GB, 2 TB€6,858.48 incl. VAT (€5,576.00 excl.); 64 GB option minus €1,273.05dell.com IE2026-11-07
ASUS Ascent GX10 GX10-GG0062BN64 GB, 1 TB€5,879, sold out (VAT status not stated)Clever, Portugal2026-11-07

Dell shows a price for the selected configuration and a delta per option; the 64 GB totals are my sums of those deltas. On both Dell pages the cart button reported a technical issue on 7 November, so I could not confirm the final checkout price. igor'sLAB reported on 28 October that the ASUS 64 GB model was open for pre-order in Germany, Austria and Switzerland with no official price. I did not find a priced 64 GB listing from HP, MSI, Gigabyte or Acer on the same day.

Two things stand out. Dell takes $2,150 off for halving the memory, a bigger gap than NVIDIA's $6,950 against $4,999. And the ASUS 64 GB unit at €5,879 costs more than the 128 GB Lenovo ThinkStation PGX with 1 TB that Bechtle Belgium listed at €5,645.86 including VAT on 24 October (see the price page; other brands are in GB10 machines compared). European 64 GB pricing has not settled, so compare a 64 GB quote with a 128 GB quote from the same vendor before deciding.

How much memory does a 64 GB Spark leave for a model?

About 53.6 GiB, by my estimate. The GB10 has one memory pool for the CPU, the GPU and the firmware, so the operating system and fixed reservations come out of the same 64 GB. I scaled from the 128 GB machine, where vLLM reports 121.6 GiB available in the logs behind the memory calculator:

128 GB unit: 128 GiB physical, 121.6 GiB seen by vLLM
             fixed reservations = 128 - 121.6 = 6.4 GiB
64 GB unit:  64 - 6.4 = 57.6 GiB seen by the GPU (assumes the same reservations)
             minus about 4 GiB for OS, CUDA context and activations
             = about 53.6 GiB for weights plus KV cache
128 GB unit: 121.6 - 4 = about 117.6 GiB

Fixed reservations hurt a small pool more. When the September 2026 kernel enabled Kexec HandOver, one owner reported 7.2 GiB less RAM at boot on a 128 GB unit (NVIDIA forum, 14 September 2026); NVIDIA's announcer in that thread attributed it to the kernel change and said mitigations were rolling out. In the other direction, the October 2026 release added Dynamic Display Reserved Memory, which shrinks the display carveout to 512 MB on a headless system (release notes), and one owner reported about 1.5 GB more available to CUDA after it (forum, 26 October 2026). On a 64 GB unit, run it headless and keep the firmware current.

Which models fit in 64 GB, with context?

MoE models up to about 35B total parameters fit at long context; 120B-class models do not fit at all. Weights are the file sizes in each Hugging Face repository; KV cache is 2 x attention layers x KV heads x head dimension x 2 bytes (BF16) per token, with the layer counts from each model's config.json. Only full-attention layers grow with context; sliding-window and DeltaNet layers hold a fixed, small state that I left out.

Model and buildWeightsKV per tokenContextWeights + KV64 GB (53.6 GiB)128 GB (117.6 GiB)
gpt-oss-20b, MXFP412.8 GiB24 KiB (12 full layers, 8 KV heads, dim 64)131,07215.8 GiBFitsFits
Gemma 4 26B-A4B, NVFP417.5 GiB20 KiB (5 global layers, 2 KV heads, dim 512)262,14422.7 GiBFitsFits
Qwen3.6-35B-A3B, UD-Q4_K_XL20.9 GiB20 KiB (10 attention layers, 2 KV heads, dim 256)262,14425.9 GiBFitsFits
Qwen3.6-35B-A3B, Q8_034.4 GiB20 KiB262,14439.4 GiBFitsFits
Qwen3.8-27B, Q8_027.0 GiB64 KiB (16 attention layers, 4 KV heads, dim 256)262,14443.0 GiBFitsFits
Qwen3.8-27B, BF16 GGUF50.9 GiB64 KiB131,07258.9 GiBNoFits
Llama 3.3 70B, AWQ 4-bit37 GiB320 KiB (80 layers, 8 KV heads, dim 128)32,768 / 131,07247 / 77 GiB32k onlyBoth
gpt-oss-120b, MXFP460.8 GiB36 KiB (18 full layers)131,07265.3 GiBNoFits
Nemotron 3.5 Super VL 120B-A12B, NVFP469.9 GiBNot computedn/aOver 64 GiB in weightsNoFits

Repository sizes read on 7 November 2026 and converted from GB to GiB. Architecture from the Gemma 4 config.json, the Qwen3.6-35B-A3B and Qwen3.8-27B cards and the gpt-oss config files. Llama weights and KV per token are the logged figures on the Llama 3.3 70B page. Gemma's sliding layers add a fixed 0.2 GiB, included. Vision projectors (about 0.9 GB each for the Qwen GGUFs) are not included. Everything here is my arithmetic, not a test on a 64 GB unit.

NVIDIA's "up to 100B parameters" for one 64 GB unit is optimistic with context in mind. At 4-bit with scales (about 0.56 bytes per parameter), 100B parameters is about 52 GiB of weights, which leaves around 1.6 GiB of my 53.6 GiB budget for KV cache. For your own model, put its numbers into the calculator with the 64 GB option, or see the KV cache size table for more models.

When is 64 GB enough?

For one person running a 20B to 35B mixture-of-experts model, 64 GB is enough, and those are the models that run fastest on a Spark anyway. Here is how I would split it:

My verdict: if your target model is already chosen and it is in the first group, the 64 GB unit saves real money. If you are buying one box to try whatever ships next quarter, buy 128 GB. Several of this autumn's notable open models are already past 64 GB at 4-bit: Nemotron 3.5 Super is 75.1 GB as NVFP4, and DeepSeek-V4-Flash has 284B parameters.

Two 64 GB units or one 128 GB unit?

One 128 GB unit, unless you specifically want the speed of tensor parallelism on a model that fits in half the memory. The pair costs more and gives you two pools, not one.

Setup for a pair is in the two-node cluster guide, and larger groups are covered in the guide to four- and eight-Spark clusters.

How do I test before buying?

Run your real model on a 128 GB Spark and measure what it uses. Load it at the context length you need, send a long prompt, and check memory with free -h (nvidia-smi reports "Memory-Usage: Not Supported" on GB10, which NVIDIA's known issues page calls expected). If the model, its KV cache and the server together stay under about 54 GiB, a 64 GB unit will hold it; if not, you know before spending $5,000. GPUwerk rents dedicated 128 GB Sparks for this at $0.79 per hour, billed per minute (how renting works). It does not rent or host 64 GB units, so the rental answers the fit question, not the 64 GB price question.

FAQ

Is the DGX Spark 64GB slower than the 128GB model?

No. NVIDIA's spec table gives one memory bandwidth for both, 273 GB/s on a 256-bit interface, and both use the same GB10 chip with up to 1 PFLOP of sparse FP4. The Register reported the bandwidth as unchanged on 2 October 2026. Decode speed depends on bandwidth, so a model that fits in 64 GB should run at the same tokens per second as on the 128 GB machine.

How much does the DGX Spark 64GB cost?

NVIDIA's blog of 2 October 2026 gives a start price of $4,999, sold only by Acer, ASUS, Dell, Gigabyte, HP and MSI. On 7 November 2026 Dell's US configurator showed $7,901 for 128 GB with 2 TB and minus $2,150 for 64 GB, which is $5,751. A Portuguese shop listed the ASUS Ascent GX10 64 GB (GX10-GG0062BN) at 5,879 euros, sold out.

Can a 64GB DGX Spark run gpt-oss-120b?

Not in a usable way. The MXFP4 weights in OpenAI's Hugging Face repository total about 65.3 GB, or 60.8 GiB, which is more than the roughly 57.6 GiB I estimate a 64 GB unit exposes to the GPU before the operating system takes its share. gpt-oss-20b, at 13.8 GB, fits with its full 131,072-token context and room to spare.

Can I cluster a 64GB Spark with a 128GB Spark?

NVIDIA's playbook for serving vLLM across Sparks says all Sparks in one cluster must have the same memory size, so a 64 GB and a 128 GB unit should not be mixed. Two 64 GB units can be linked with one QSFP cable; NVIDIA rates that pair at models up to 200B parameters and reports up to 1.7x the speed of one unit on Qwen3.8 27B.

Is 64 GB enough for a 70B model?

For one user at short context, yes. A 70B model at 4-bit AWQ is about 37 GiB of weights, and Llama 3.3 70B adds 320 KiB of KV cache per token, so 32,768 tokens comes to about 47 GiB in total by my arithmetic. At 65,536 tokens it passes the roughly 53.6 GiB I estimate is usable. Speed is the same 6 tok/s class as on 128 GB.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

First top-up: pay $20, get $30 in credit

Check your model against 64 GB on a real Spark.

Rent a 128 GB Spark by the minute, load your model at your real context length, and see whether it stays under the 54 GiB a 64 GB unit would give you.

A dedicated DGX Spark in the EU, $0.79/hour, billed per minute. $30 of credit is about 37 hours of compute.

Create account Open the memory calculator