The DGX Spark 64GB goes on sale today, October 23, at $4,999. That is half the memory of NVIDIA’s original desk-sized AI box, for more money than the 128GB model launched at.
It only makes sense if your models fit in 64GB. If they don’t, you are paying more for less, and you can’t add memory later.
Here are the numbers that matter before you read on:
- Price: starts at $4,999, per TweakTown
- Release date: October 23, 2026
- Sold only through partners: Acer, ASUS, Dell, GIGABYTE, HP and MSI
- Memory: 64GB, with a claimed ceiling of about 100 billion parameters
- Same chip, same 273 GB/s memory bandwidth as the 128GB model
What the DGX Spark 64GB changes (and what it doesn’t)
NVIDIA kept almost everything the same. The chip, the 20-core Arm CPU, the 200 Gbps networking and the software stack carry over unchanged.
The one real difference is memory. You get 64GB of unified LPDDR5X instead of 128GB, so the largest model you can load shrinks by half.
| Spec | DGX Spark 64GB | DGX Spark 128GB |
|---|---|---|
| Unified memory | 64GB | 128GB |
| Claimed model size | Up to ~100B parameters | Up to ~200B parameters |
| Memory bandwidth | 273 GB/s | 273 GB/s |
| Chip | GB10 Grace Blackwell | GB10 Grace Blackwell |
| Networking | ConnectX-7, 200 Gbps | ConnectX-7, 200 Gbps |
| Sold by | Partners only | Founders Edition and partners |
Specs here come from Petronella Tech and TweakTown. Note that Petronella sells accessories for the box, so treat its benchmarks as a seller’s view.
Why the $4,999 price stings
The original 128GB model was announced at $3,999. NVIDIA’s own Founders Edition later moved to $4,699, so the 64GB box starts $300 above that.
Partner pricing can also drift. Petronella reports 128GB street prices above $6,000, and says the $4,999 figure may not hold. Memory costs are part of the story, and we covered how Micron’s strike is pushing RAM prices up earlier this month.
There is also cheaper competition. TweakTown points to AMD Ryzen AI Max+ 395 mini PCs with 128GB, with one GMKtec model listed at $3,650. They run a different software stack, so the comparison isn’t one-to-one. If you already live in NVIDIA’s CUDA tools, the Spark saves setup time. If you mostly run open models through common runtimes, the price gap matters more.
What you can actually run on 64GB
For most local AI work, 64GB is plenty. NVIDIA lists day-zero support for Qwen3.8 27B and Gemma 4 26B-A4B, both of which fit with room to spare, according to the NVIDIA announcement.
Recursive’s new Flint coding model, a 40B mixture-of-experts with only 2.5B parameters active, is another good fit for this class of machine.
A rough rule helps here. Model weights, the context cache and the operating system all share the same 64GB pool. A 4-bit 70B model can eat most of it before you add a long prompt, so leave headroom rather than aiming for the claimed maximum.
Where you will feel the squeeze is at the top end. TweakTown expects models around 80B parameters and above, and memory-hungry fine-tuning, to be limited.
Pairing two units to reach 128GB
You can link two boxes with a single 0.5 m QSFP112 cable, port 0 to port 0. Pooled memory then reaches 128GB, with a claimed ceiling near 200B parameters.
Two units cost about $10,000 before the cable, so this only helps if you already own one. The link is also slow compared with local memory: roughly 25 GB/s against 273 GB/s on each box. NVIDIA claims up to 1.7x performance on a Qwen3.8 27B test, which is a vendor number.
If you want the Windows side of NVIDIA’s local-AI push, our RTX Spark release date and price guide covers the PC chips.
Who should buy it, and who should skip it
The right answer depends on the size of the models you run and how soon you need the machine.
| If you… | Our read |
|---|---|
| Run 20B to 70B models for coding, chat or agents | A good fit. 64GB gives you room for the model plus a long context window. |
| Fine-tune larger models or test 100B+ weights | Skip it. Memory is the limit, and you cannot upgrade it later. |
| Already own a 128GB Spark | Skip it, unless you want a cheap second node for clustering. |
| Want the lowest price per gigabyte | Look at 128GB AMD mini PCs, but expect a different software setup. |
NVIDIA says it is keeping the platform at an accessible price point, per its announcement. With the base price now above the old launch price, that claim is debatable.
Frequently Asked Questions
How much does the DGX Spark 64GB cost?
It starts at $4,999 through NVIDIA’s partners. Final prices can vary by maker and configuration.
When can you buy the DGX Spark 64GB?
Sales start October 23, 2026. It is sold only through Acer, ASUS, Dell, GIGABYTE, HP and MSI, not directly by NVIDIA.
Is 64GB enough for local LLMs?
Yes for models up to roughly 100B parameters, which covers most popular open models. Larger models and heavy fine-tuning are where it runs short.
Can you upgrade a 64GB DGX Spark to 128GB?
No, the memory is built in. You can link two units to pool 128GB, but that means buying a second box.
Our take: buy the DGX Spark 64GB only if you know your models fit and you want NVIDIA’s software stack. If you are not sure, a cheaper 128GB mini PC or waiting for prices to settle is the safer bet, because $4,999 for half the memory is a hard sell today.


Leave a Reply