DGX Spark 64GB vs 128GB: Which to Buy After the October Price Split

Short answer: buy the 64GB DGX Spark only if the model you will load fits in 64GB of advertised unified memory and you want the CUDA appliance. Buy the 128GB unit if the checkpoint is in the class a 64GB pool cannot hold. Skip both if your target is Qwen 3.8 27B Q4. One used RTX 3090 still costs less for that file: CompSniper’s sold median was $1,499 on 6 October 2026 (typical range $1,329–$1,599), against a 64GB Spark starting at $4,999. Skip both if you do not need CUDA (a Strix Halo mini PC or a Mac). The 64GB box is not for sale yet.

NVIDIA’s blog on 2 October 2026 says the 64GB configuration ships from Acer, ASUS, Dell, Gigabyte, HP, and MSI on Friday 23 October 2026, starting at $4,999, with the same GB10 Grace Blackwell Superchip, DGX OS, and NVIDIA software stack as the 128GB model (NVIDIA blog). The Register the same day reported the 128GB list at $6,950. Today is 6 October 2026. The NVIDIA product page fetched that day still labeled the 64GB unit Coming Soon.

What changed, and what is the same chip?

The memory configuration and the price split. The GB10, the 273 GB/s figure, and the ConnectX-7 port are the shared platform.

Fact	64GB OEM Spark	128GB DGX Spark	Where the number comes from
Unified memory	64GB LPDDR5X	128GB LPDDR5x	NVIDIA product page, fetched 2026-10-06
Memory bandwidth	273 GB/s on the shared spec table	273 GB/s	Same table. The Register says bandwidth is unchanged
Vendor parameter ceiling, one box	Up to 100B	Up to 200B	NVIDIA product-page cluster matrix, fetched 2026-10-06
Ship status on 2026-10-06	Coming Soon. Partners from 2026-10-23	Already the existing SKU	Product page + NVIDIA blog
Price	Starting at $4,999	$6,950	NVIDIA blog (starting price). The Register (128GB list, 2026-10-02)
Storage	The Register says half the storage. No terabyte figure in that report	Spec table says up to 4TB NVMe, and does not split the number by SKU	Leave the 64GB drive size empty
Power	Shared table: 240W supply, GB10 TDP 140W	Same table, not a per-SKU measurement	NVIDIA product page, fetched 2026-10-06
LocalRig tok/s	Not yet measured	Not a new first-party row	—

The Register also says the 64GB systems keep the same 20-core Arm CPU and ConnectX-7, and that clustering still goes up to four GB10 machines at 200 Gbps. The product-page spec table lists that CPU as 10 Cortex-X925 plus 10 Cortex-A725, and the NIC as ConnectX-7 at 200 Gbps. Those are platform specs. They are not a LocalRig teardown of a 64GB unit, because none was shipping on the day this page was written.

LocalRig’s hardware catalog had recorded a $3,999 retailer observation for the 128GB Spark on 2026-07-10. That row is now a historical observation, not the October list. Some older roundups still say $4,699. This page does not treat $4,699 as the current 128GB price. The dated figure for 128GB is The Register’s $6,950 on 2 October 2026. NVIDIA’s own page, fetched 6 October 2026, showed a Buy button and did not expose a dollar amount in the page text, so this is a journalism report, not a LocalRig receipt.

Two 64GB units at the published starting price are 2 × $4,999 = $9,998 before tax, a cable, or a partner price above the floor. That arithmetic is not a bundle SKU. One 128GB unit at the reported $6,950 is the smaller bill for a single 128GB pool.

Which models fit in 64GB versus 128GB?

Fit the weight file first. NVIDIA does not publish a usable-after-OS number, and LocalRig has not measured one.

The product page states coherent unified memory of 64GB or 128GB. It does not say how many of those gigabytes remain after DGX OS. ComputingForGeeks (Josphat Mutai, updated 3 October 2026) estimates about 120GB usable on the 128GB Spark and says there is no fixed split. That is their estimate. They do not publish a measured usable figure for the 64GB SKU. Where a weight file sits close to the advertised pool, the cell below says not yet measured instead of a guessed “yes.”

Use the VRAM calculator for a context-length pass on a dense model. It is an estimator. It is not a Spark measurement.

Workload	Cited size	64GB advertised pool	128GB advertised pool
Qwen3.8-27B, Unsloth UD-Q4_K_XL	17.6GB file. Unsloth’s 4-bit band is 16–19GB before long context (card and docs, accessed 2026-09-06)	Weights fit. Long-context KV on this SKU: not yet measured	Weights fit. One used 3090 is still the lower CUDA bill at CompSniper’s $1,499 sold median (6 October 2026). See Can I run Qwen 3.8 locally?
Qwen3.8-27B Q8_0	29GB file (same Unsloth card)	Weights fit. OS + KV headroom: not yet measured	Weights fit
Qwen3.8-27B BF16	54.7GB file (same card)	The file is under 64GB. Room for the OS and a context window: not yet measured	Weights fit
70B Q4 weights	LocalRig memory math: about 35–40GB for weights before KV cache	Inside NVIDIA’s vendor ceiling of 100B parameters. A comfortable KV total on 64GB: not yet measured	NVIDIA’s fine-tune sentence is written against 128GB, up to 70B parameters. Q4 inference weights fit the advertised pool. See 70B hardware
Qwen3.8 Flash-Next GGUF, Unsloth 1-bit	About 75GB of RAM, with a 96GB machine preferred (Unsloth, accessed 2026-09-07)	No	The weight size is under 128GB. Context on top of that depends on usable memory, which LocalRig has not measured. See Flash GGUF
Qwen3.8-Flash-Next NVFP4 (Mia)	About 124GiB on disk. The dual README wants about 126GiB free on each node (accessed 2026-09-07)	No	Mia’s single-Spark runbook is a 128GB GB10, not a 64GB box. See dual Spark how-to
GLM-5.3-Flash, Unsloth 1-bit	About 100GB total memory. The 3-bit band is 128–150GB (Unsloth, accessed 2026-09-10)	No	1-bit is the candidate on a 128GB box. 3-bit is at or above the advertised pool. See GLM-5.3-Flash
gpt-oss-20b MXFP4	OpenAI: runs within 16GB	Fits, and is the wrong reason to buy either Spark	Same
Vendor ceiling	Not a file size	Up to 100B parameters (NVIDIA)	Up to 200B parameters (NVIDIA)

Read the vendor ceiling as a vendor ceiling. “Up to 100 billion parameters” does not say the quant, the context length, or the runtime. A 54.7GB BF16 file can be under 64GB on disk and still fail once the OS and the KV cache are in the same pool. That failure mode is not yet measured here.

Decode speed is a separate question, and it is empty for the new SKU. NVIDIA lists 273 GB/s for the platform, and The Register says the 64GB units keep that bandwidth. A model that fits in both pools should not be assumed to run at half speed on the 64GB box. LocalRig has not measured that. Do not paste a 128GB community tok/s number onto the 64GB SKU.

Who should buy the 64GB, the 128GB, or two 64GB units?

Match the pool to the checkpoint. Then decide whether CUDA is worth the premium.

Buy the 64GB OEM Spark when

• The weight file, plus the context you will actually use, fits in 64GB. Qwen3.8-27B at Unsloth 4-bit is in that class on file size. A 70B Q4 weight file is in NVIDIA’s vendor range and is not yet measured for long context on this SKU.
• You want DGX OS, CUDA, vLLM, and ConnectX-7 in one small chassis, and you do not need the 75GB-class or 124GiB-class checkpoints.
• You can wait until 23 October 2026. Buying the existing 128GB listing early does not get you this SKU.
• The starting price of $4,999 is acceptable for that smaller pool. Partners can charge more. LocalRig has no street price above the NVIDIA floor.

NVIDIA’s blog says one 64GB system supports up to 100B-parameter models, and names Qwen 3.8 27B as the model its new launcher will download. The Register says NVIDIA is pointing this SKU at private agents in the 26–35B class, with Qwen 3.8 27B as the example, and that the smaller pool is less suited to fine-tuning. The product page’s “fine-tune models up to 70 billion parameters” sentence is attached to 128GB of unified memory. Do not move that sentence onto the 64GB box.

There is no Amazon link on this page for the 64GB unit. NVIDIA says it is exclusive to those OEM partners. The existing DGX Spark Amazon listing is the 128GB founders path LocalRig verified as ASIN B0FWJ16CCH on 2026-07-10. Read the title and the live price before you treat that listing as the October 128GB unit. It is not the 64GB SKU.

Buy the 128GB Spark when

• The checkpoint does not fit in 64GB. Unsloth’s ~75GB Flash GGUF and Mia’s ~124GiB NVFP4 stack are the concrete cases already documented on LocalRig.
• You want one memory pool, not two machines clustered in software.
• You accept The Register’s $6,950 report as the list to beat, and you re-check the seller on the day you pay. ComputingForGeeks (updated 3 October 2026) says that author checked vendor stores and Amazon that week and also lists the 128GB Spark at $6,950. That is a second dated report, not a LocalRig checkout.

Ownership against a dual-3090 workstation — noise, power, 128GB unified versus 48GB split — stays on DGX Spark vs 2× RTX 3090. That page is about the 128GB appliance. It is not a 64GB review.

Buy two 64GB units only when you already want a cluster

NVIDIA’s blog says two 64GB systems, linked with a QSFP cable and set up by NVIDIA Sync Cluster Assistant, pool memory to 128GB, support up to 200B parameters, deliver twice the memory bandwidth, and in NVIDIA’s Qwen 3.8 27B test delivered up to 1.7× the performance of one system. The product-page matrix matches that shape: 2× DGX Spark / 128GB / up to 200B, then a separate row for 2× / 256GB / up to 400B. The 256GB row is two 128GB machines, not two 64GB machines.

The 1.7× figure is a vendor result. The blog does not publish the runtime version, quantization, context length, batch size, or driver. LocalRig will not turn it into a tok/s. Sync Cluster Assistant and the end-of-month model launcher are software NVIDIA says is coming. They are not a measured LocalRig install.

Two 64GB boxes also do not run Mia’s dual-Spark Flash-Next recipe. That runbook wants two 128GB nodes and about 126GiB free on each for the NVFP4 checkpoint. Pooled 128GB across two smaller machines is a different topology, and each 64GB node is too small for that file.

At starting prices, two 64GB units ($9,998) cost more than one reported 128GB unit ($6,950) and give you a second GB10 plus a second 273 GB/s pool. Buy the pair if you want that second chip. Buy one 128GB if you want one pool and a lower bill.

When is a used 3090, a Strix Halo box, or a Mac the better buy?

When CUDA-in-a-tiny-NVIDIA-box is not the constraint you are paying for.

Used RTX 3090, for Qwen 3.8 27B Q4 and anything else that fits in 24GB. Unsloth’s 4-bit 27B band is 16–19GB. A 24GB card leaves KV room that a 16GB card does not.

CompSniper, fetched 6 October 2026, says a RTX 3090 sells for a median of $1,499 across the 204 most recent completed eBay sales, with most in $1,329 to $1,599. The most recent sale on that page is 6 October 2026. The same summary lists an average of $1,458, about $19 shipping, and extremes of $580 and $3,650. Those extremes are not the typical band. LocalRig did not recompute the median. The visible sample includes at least one RTX 3090 Ti row, so this is the tracker’s published figure, not a 3090-only audit.

RigPrice, same fetch, is asking prices. Its 2026-10-06 row lists a used going rate of $1,655 (median ask) across 92 active used listings, with the middle half of asks at $1,550–$1,843. The 2026-10-04 row was $1,665. The page says 24 used listings delisted that day, typically at $1,500, and it does not treat a delisting as a confirmed sale.

LocalRig’s $580–$780 band in prices.json, dated 2026-08-01, is stale. The October sold median sits near twice the top of that band. Do not quote it as this week’s price. Card condition, cooler, and seller still belong on the used 3090 buying guide.

The 27B Q4 verdict still holds. A $1,499 sold median, and a $1,655 asking rate, are both under a 64GB Spark starting at $4,999 and under the reported $6,950 128GB list. Skip both Sparks when that 17.6GB file is the job. What changed is the gap: this is no longer a sub-$800 card.

Two cards are a different sum. 2 × the $1,499 median is $2,998 before shipping and before a PSU, board, and case. That is arithmetic on the per-card median, not a paired sold listing. $2,998 in cards is still under both Spark lists. It is the 48GB-split path for a 70B Q4, and it is still not a 128GB unified pool. The build notes are the dual 3090 guide.

Browse used RTX 3090 24GB on eBay →

Strix Halo mini PC, when you want a lot of unified memory and you do not need CUDA. ComputingForGeeks (updated 3 October 2026) reports a Bosgame M5 128GB at $3,324.04 and a GMKtec EVO-X2 128GB at $3,649.99, and says those prices were checked on Amazon that day. LocalRig did not re-open those listings on 6 October 2026. Those boxes are ROCm and Vulkan, not DGX OS. LocalRig’s older best mini PC prices predate this October split. The platform comparison that is still the right frame is Strix Halo vs Mac Studio.

Strix Halo 128GB mini PC search →

Mac, when one person wants decode bandwidth and can live on MLX or Metal. The same ComputingForGeeks article reports a Mac Studio M5 Max, 40-core GPU, 128GB, 512GB SSD, at $5,099, and says that figure came from Apple’s configurator on 3 October 2026. LocalRig did not re-open the configurator on 6 October 2026. That reported price sits under The Register’s $6,950 Spark list and above the Strix prices in the same roundup. It is not a CUDA machine. Memory-class guidance that is not an M5 configurator check: how much unified memory.

Mac Studio 128GB on Amazon →

Rent, when you have not proved the model is worth a $5,000–$7,000 box. Run the rent vs buy break-even before you pay for a Spark. A short rental on RunPod or Vast.ai is the way to learn that a 27B Q4 was the real job. Neither marketplace is a DGX Spark rental. Do not treat an August 2026 hourly quote on LocalRig as the October rate.

What this page will not fill in

• 64GB tok/s. Not yet measured. No named public packet with URL, date, model, quant, runtime, and a 64GB Spark existed for this draft. The unit was not shipping.
• Usable memory after DGX OS on either SKU, as a LocalRig measurement. The 120GB figure is ComputingForGeeks’ estimate for the 128GB box only.
• 64GB drive capacity in terabytes. The Register says half the storage. NVIDIA’s spec table says up to 4TB and does not split it by SKU.
• A partner’s price above $4,999. Unknown until those stores list the SKU.
• An NVIDIA cart screenshot of $6,950. The Register and ComputingForGeeks report it. This page did not complete a purchase.

Who this is NOT for

• Anyone whose only model is Qwen 3.8 27B at Q4. One used 3090 is still the lower CUDA bill at CompSniper’s $1,499 sold median (6 October 2026), versus $4,999 to start on the 64GB Spark. Spark’s price, at either SKU, is the wrong spend for that file.
• People who need Mia’s Flash-Next NVFP4 stack on two nodes. That is two 128GB Sparks, documented in the dual-Spark how-to. Two 64GB units do not hold a ~124GiB checkpoint on each node.
• Buyers who wanted the 2025 launch price. The Register puts the 128GB list at $6,950 on 2 October 2026, and the 64GB starting price at about $4,999, which is still above what the 128GB sold for at launch. Waiting for $3,999 to return is a different decision. This page has no evidence that it will.
• Shoppers who do not need CUDA. A Strix Halo box or a Mac can hold a large unified-memory model without DGX OS. Pay the Spark premium for the NVIDIA stack, not for a logo.
• Anyone looking for a LocalRig speed winner between 64GB and 128GB. There is no first-party row. The 1.7× cluster claim is NVIDIA’s, and it is not a single-box comparison.

Sources

• NVIDIA blog, 2026-10-02: 64GB from Acer, ASUS, Dell, Gigabyte, HP, and MSI on 23 October 2026, starting at $4,999. Same GB10 stack. Vendor claims: up to 100B parameters on one 64GB unit; two units pool to 128GB and up to 200B; up to 1.7× on Qwen 3.8 27B versus one system. Accessed 2026-10-06.
• NVIDIA DGX Spark product page, fetched 2026-10-06: 64GB Coming Soon via OEMs; 64GB or 128GB coherent unified memory; 273 GB/s; up to 4TB NVMe on the shared table; ConnectX-7 at 200 Gbps; 240W supply; GB10 TDP 140W; cluster matrix from 1× 64GB / 100B through 4× / 512GB / 700B. Fine-tune-up-to-70B sentence is attached to 128GB.
• The Register, Tobias Mann, 2026-10-02: 128GB list reported at $6,950; 64GB described as half the memory and half the storage, expected around $4,999; bandwidth reported unchanged at 273 GB/s.
• ComputingForGeeks, Josphat Mutai, updated 2026-10-03: third-party roundup. Spark 128GB at $6,950; 64GB from $4,999; Strix and Mac prices cited above as that author’s 3 October store checks. Usable-memory estimate of about 120GB on the 128GB Spark is theirs. LocalRig did not re-check those carts on 2026-10-06, and did not adopt that article’s tok/s table.
• Unsloth Qwen3.8-27B GGUF file sizes and 4-bit band, accessed 2026-09-06: Hugging Face, docs.
• Unsloth Flash-Next GGUF note, accessed 2026-09-07: docs.
• MiaAI-Lab Spark runbooks, accessed 2026-09-07: single, dual.
• Unsloth GLM-5.3-Flash memory bands, accessed 2026-09-10: docs.
• LocalRig catalog observation of $3,999 on 2026-07-10, retained as history in verticals/hardware/parts.yaml. The 1 August 2026 used-3090 band of $580–$780 in site/src/data/prices.json is superseded for this page by the October trackers below.
• CompSniper, RTX 3090 worth, fetched 2026-10-06: median $1,499 across 204 completed eBay sales, typical range $1,329–$1,599, average $1,458, about $19 shipping, most recent sale 2026-10-06. Visible sample includes an RTX 3090 Ti row. LocalRig did not recompute the median.
• RigPrice, RTX 3090, fetched 2026-10-06: used asking going rate $1,655 on the 2026-10-06 row (92 active used listings, middle half $1,550–$1,843). The 2026-10-04 row was $1,665. Asking prices, not sold prices.