NVIDIA halves DGX Spark's memory to 64GB, and the entry price still lands at $4,999

NVIDIA announced on October 2 a 64GB version of DGX Spark, its desktop machine for running and developing AI models locally. It goes on sale Friday, October 23, starting at $4,999, and only through six hardware partners: Acer, ASUS, Dell, Gigabyte, HP and MSI. NVIDIA will not sell this configuration itself; its product page says the 64GB model is available exclusively through participating OEMs.

Everything else is unchanged from the existing 128GB model, according to NVIDIA: the same GB10 Grace Blackwell chip, the same DGX OS and software stack, the same built-in ConnectX-7 network card. What changes is half the unified memory, and with it the size of model the box can hold.

What you get, and what NVIDIA claims

• Model size: NVIDIA says a single 64GB unit runs models of up to 100 billion parameters on device. Two units connected with a QSFP cable pool their memory to 128GB and, per NVIDIA, handle models of up to 200 billion parameters.
• Clustering: a new NVIDIA Sync Cluster Assistant detects the second unit and configures the network. In NVIDIA's own test with Qwen 3.8 27B, two clustered 64GB systems were up to 1.7x faster than one. That figure is the vendor's; no independent measurement is available yet.
• Software: NVIDIA lists llama.cpp, Ollama, vLLM, LM Studio and PyTorch as supported. An NVIDIA Sync Model Launcher, due at the end of October, is meant to download and start Qwen3.8 27B with a few clicks and wire it into the OpenCode coding agent.
• Specs that did not change: the published spec sheet still lists 273 GB/s of memory bandwidth and up to 1 petaFLOP of FP4 compute for both memory sizes.

The price is the awkward part

The 64GB unit is a cheaper entry point only relative to today's lineup. When the original 128GB DGX Spark reached reviewers in October 2025, The Decoder reported a price of about $4,000. A year later, half the memory starts at roughly $1,000 more. Some outlets reported on October 2 that the 128GB model's price has also gone up; Promptea could not open those reports or confirm a current NVIDIA list price, so we are not repeating a figure. NVIDIA's announcement does not explain the pricing.

NVIDIA also notes that $4,999 is a starting price, so the final cost depends on how each partner configures its version.

What 64GB actually buys for local models

The 100-billion-parameter ceiling is a best case. NVIDIA's post does not state the precision behind it. As a rough calculation of our own: at 4-bit precision, the weights of a 100B model take about 50GB, and that memory is shared with the operating system and everything else on the machine. What is left over is what the model has for its KV cache, which is what grows as the context window fills. In practice, a model near the ceiling leaves little room for long prompts, large documents or several agents running at once; NVIDIA's own use-case list points to a second unit for "longer context windows or multiple agents."

Speed is a separate limit. Generating each token requires reading the model's weights from memory, so bandwidth, not capacity, sets the ceiling on output speed. The 64GB model keeps the same 273 GB/s as the 128GB one. Reviews of the original Spark in 2025 described it as built for capacity rather than raw speed, and nothing in this announcement changes that.

Who this is for

For developers who want a coding or research agent running around the clock on their own hardware, mid-size open models in the 20B to 30B range are the realistic target on a single unit, and they are the ones NVIDIA itself is showcasing. The pitch is privacy and no per-token billing. The trade-off is a $5,000 upfront cost, output speeds bounded by memory bandwidth, and open models that may trail the frontier APIs on hard tasks.

The clustering path is the more interesting idea: start with one 64GB box and add a second when the workload grows. But two 64GB units together add up to the same memory as one 128GB unit, at a combined starting price of nearly $10,000, so buyers who already know they need 128GB should compare that against current 128GB pricing before choosing the modular route.

All performance and capacity figures in this article come from NVIDIA. Independent reviews of the 64GB configuration are not yet available.