NVIDIA ships a $4,999 64GB DGX Spark for local AI amid the memory crunch

NVIDIA on Friday said DGX Spark will be available with 64GB of unified memory from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Oct. 23 at $4,999, keeping the GB10 Grace Blackwell Superchip and full NVIDIA AI stack while offering a cheaper on-ramp than the 128GB model.

Memory prices are the quiet tax on local agents. Cutting DGX Spark to 64GB at $4,999 — with a clean path to cluster two into 128GB — is NVIDIA admitting the 128GB SKU priced a lot of builders out, and selling the on-ramp without abandoning the Grace Blackwell stack.

On Friday, 2 October 2026, NVIDIA said a DGX Spark with 64GB of unified memory is coming from partner computer makers. The company blog is titled “NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI.” The byline is Allen Bourgoyne. The page dates the post October 2, 2026, and does not print an hour. The line under the headline says that starting Friday, Oct. 23, the new configuration will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI, and that NVIDIA Sync Cluster Assistant scales projects and workloads. Those lines are NVIDIA’s.

What the price is, and who sells it. The blog says DGX Spark 64GB is available from those six companies on Friday, Oct. 23, starting at $4,999. Starting at is the opening price NVIDIA prints. It is not a line that every partner’s box will be that number on the day. Unified memory means the processor and the graphics chip share one pool of memory, instead of each keeping its own. Sixty-four gigabytes is the new size. The same post calls the existing machine the 128GB model, twice the memory. The product page says the 64GB configuration is coming soon, and only through participating partners. Those lines are NVIDIA’s.

What the cheaper machine keeps. The blog says the 64GB configuration keeps the GB10 Grace Blackwell Superchip, DGX OS, and the full NVIDIA AI software stack, the same as the 128GB model. The GB10 is NVIDIA’s chip that puts a Grace processor and a Blackwell graphics processor on one package. DGX OS is the operating system that ships on the box. The software stack is the set of NVIDIA programs that come with it. The post also says the new configuration is available only from the manufacturer partners, not as a separate NVIDIA-sold box the blog names. Those lines are NVIDIA’s.

What one machine can run, as NVIDIA states it. A single 64GB unit supports up to 100-billion-parameter models, and the agent programs built on them, fully on the device. A parameter is one number the model learned while it was trained. One hundred billion of them is a large model. An agent, here, is software that takes a next step, not only a chat reply. Fully on the device means the model runs on the machine itself. The blog says that can be private, without depending on a cloud service for every answer. Up to is NVIDIA’s ceiling. The product page’s scale table prints the same row: one DGX Spark at 64GB, up to 100 billion parameters. Those lines are NVIDIA’s.

What two machines do together. The blog says two units can cluster through NVIDIA Sync Cluster Assistant, and that this does not need extra setup. A cluster means the two machines work as one. Two 64GB units pool their memory to 128GB and expand support to up to 200 billion parameters. Pool means the software treats the two memory banks as one larger bank. The same paragraph says the link delivers twice the memory bandwidth and up to 1.7 times the performance. Memory bandwidth is how fast the chip can read and write that memory. The product page has a matching row: two DGX Spark systems whose combined memory is 128GB, up to 200 billion parameters. The blog says the two machines do not merely double the memory. Those figures are NVIDIA’s.

The test behind the 1.7 times. In NVIDIA’s Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7 times the performance of one system. Qwen 3.8 27B is an open model with 27 billion parameters. The blog writes the name as Qwen 3.8 27B and as Qwen3.8 27B. Up to 1.7 times is the result NVIDIA reports for that test. The post does not print the same multiple for every model, and it does not print the test’s settings beyond the model name. That result is NVIDIA’s.

How the two machines are plugged together. Every DGX Spark ships with a built-in ConnectX-7 network card. A network card is the port that lets one computer talk to another at high speed. Two units can connect directly with a QSFP cable, the plug used on that kind of link. NVIDIA Sync finds the connected units, checks that they are configured, and sets up the ConnectX-7 network. The blog says each machine already runs the same software, so a second unit does not mean rebuilding the setup. One example on the post says the two 64GB systems connect over a 200-gigabit Ethernet link. That is the speed NVIDIA names for the link. The product page says ConnectX networking can join up to four DGX Spark systems, for larger models, faster answers, and several agents at once. Four is the product page’s ceiling. Friday’s blog walks through two. Those lines are NVIDIA’s.

What is on the machine the day it arrives. The blog says it ships ready for agent work: NVIDIA Agent Toolkit, the CUDA-X AI libraries, Nemotron open models, and programs including Ollama, vLLM, and PyTorch with CUDA. CUDA is NVIDIA’s software layer that lets those programs use the chip. A person loads a model in one of those programs and runs it locally. The getting-started list on the same post also names llama.cpp and LM Studio, two more programs people already use to run models on their own hardware. The blog says a developer can go from turning the machine on to a running model in minutes. Those names are NVIDIA’s.

What NVIDIA says is still on the way. Blender, the 3D program, is among the first big creator applications to support the platform, and a prebuilt installer is coming soon. At the end of the month, NVIDIA Sync Model Launcher is supposed to download and start Qwen 3.8 27B on one DGX Spark or on a cluster, set the model up across the connected machines, and open it from a laptop. The launcher will also set up OpenCode so a person can code in a browser. Coming soon, and the end of the month, are the blog’s timing. The day-one list above is the stack the post says is supported out of the box. The product page rates the Grace Blackwell design at up to 1 petaFLOP of AI performance at FP4. A petaFLOP is a thousand trillion arithmetic steps a second. FP4 is a short number format that lets the chip do more of those steps on AI work. That rating is the product page’s, for the chip the blog says the 64GB machine keeps.

Which product-page rows belong to the bigger machines. The page’s fine-tuning line is written next to 128GB of unified memory: fine-tune models up to 70 billion parameters. Fine-tuning means teaching an existing model on your own examples. The blog does not print that 70 billion figure for the 64GB unit. The scale table also lists two 128GB systems at 256GB combined and up to 400 billion parameters, and four systems at 512GB combined and up to 700 billion parameters. Those rows are the product page’s description of larger clusters. They are not the capacity of one 64GB box.

What the other papers reported the same day. Tom’s Hardware and The Register both carried the announcement: the 64GB configuration, the same six partners, Oct. 23, and a starting price of $4,999. Tom’s Hardware also says the 20-core Arm processor and 273 gigabytes per second of shared memory bandwidth carry over from the original Spark. The Register says the new system has half the memory and half the storage, and that the 128GB version’s price rose to $6,950. A core is one processor inside the chip. Those extra figures are the papers’. The NVIDIA blog does not print a core count, the 273 gigabytes per second, a storage size, or a new price for the 128GB model.

The picture is a product photograph. Seven compact desktop units sit in a row on a black field, edged in green light. The marks on the machines are Acer, ASUS, Dell, Gigabyte, HP, Lenovo, and MSI. The blog’s sellers for the 64GB DGX Spark are Acer, ASUS, Dell, Gigabyte, HP, and MSI. The same post says Windows PCs powered by RTX Spark are coming this month from Acer, ASUS, Dell, HP, Lenovo, Microsoft, and MSI. Lenovo is on that RTX Spark list, and on the photograph. The blog does not name Lenovo as a seller of the 64GB DGX Spark. The frame does not print a price or a calendar date.

In plain terms, NVIDIA said on Friday that partner makers will sell a DGX Spark with 64GB of shared memory from Oct. 23, starting at $4,999. The machine keeps the GB10 Grace Blackwell chip, DGX OS, the ConnectX-7 network port, and the NVIDIA software stack of the 128GB model. NVIDIA says one unit runs models up to 100 billion parameters on the device. Two units, joined by a cable and set up by NVIDIA Sync, pool to 128GB and up to 200 billion parameters. In NVIDIA’s Qwen 3.8 27B test, the pair was up to 1.7 times as fast as one machine. Tom’s Hardware and The Register reported the same price, date, and partners. The core count, the memory-speed figure, the storage cut, and the $6,950 line for the 128GB model are theirs, not numbers the NVIDIA blog prints.