NVIDIA announces 64GB DGX Spark for October 23 at a $4,999 starting price

NVIDIA announced a 64GB unified-memory version of its DGX Spark local AI computer on October 2. The company says it will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI on October 23, starting at $4,999. For developers weighing a machine for on-device AI work, the new configuration offers a lower-memory entry to the DGX Spark platform; its advertised capacity and performance figures come from NVIDIA, not independent tests of the new model.

What changes in the 64GB DGX Spark

The new configuration has half the unified memory of the existing 128GB DGX Spark. NVIDIA says it retains the GB10 Grace Blackwell Superchip, DGX OS and the company’s AI software stack. It is being offered through manufacturer partners, rather than as a replacement for the 128GB model. The announced $4,999 figure is a starting price in US dollars; NVIDIA’s announcement does not specify regional prices, taxes or the final configurations each partner will sell.

NVIDIA says a single 64GB system can run models with up to 100 billion parameters entirely on the device. That upper limit is a company specification, not a promise that every model of that size will run equally well. The announcement does not spell out the model formats or workload settings behind the figure, information a buyer would need to judge whether a particular model and context length will fit.

The software pitch covers more than model inference. NVIDIA says DGX Spark ships with its Agent Toolkit, CUDA-X AI libraries and Nemotron open models, and supports Ollama, vLLM and PyTorch with CUDA. The company describes local operation as a way to use developers’ own data without relying on a cloud instance for every task. Its privacy and cloud-independence claims describe that intended use; the cited material does not independently establish a privacy outcome for a particular application.

How NVIDIA says two systems can work together

In a test using Qwen 3.8 27B, NVIDIA says two clustered 64GB systems delivered up to 1.7 times the performance of one. The figure is the company’s result for that test, not a general speedup or an independently reproduced benchmark. NVIDIA also says a Sync Model Launcher is planned for the end of October. It is intended to launch Qwen3.8 27B on one DGX Spark or a cluster and make the model accessible from a laptop; that feature remains a planned release.

What independent testing says about DGX Spark

Tom’s Hardware reviewed the existing 128GB DGX Spark in January, months before this 64GB announcement. Its review described the platform as a capable local-AI toolkit with CUDA support, shared memory and ConnectX-7 networking. It also questioned whether the high price would make sense for buyers who would not use its features extensively. Those observations provide context for the platform, but the review did not test the newly announced configuration or NVIDIA’s two-unit 64GB performance claim.

The memory distinction matters for local AI work. Tom’s Hardware noted that large models demand substantial memory and that longer contexts add to the pressure beyond the model’s own size. It contrasted DGX Spark’s shared 128GB memory with the RTX 5090’s 32GB of onboard memory. The review also pointed to AMD Ryzen AI Max+ 395 systems as another route to a large unified-memory pool, while noting their lack of native CUDA support. Those comparisons concern the broader choice of hardware; they do not establish how the 64GB DGX Spark performs.

Availability and unanswered questions

NVIDIA names six manufacturers and an October 23 availability date, but its announcement does not identify retailer stock, country-by-country availability or each partner’s final configuration. Independent testing of the 64GB system is also still needed to assess model fit, real workload speed and the benefit of clustering. For now, buyers can compare the stated hardware, software support and starting price while treating the parameter limits and 1.7-times result as NVIDIA’s claims.