Nvidia's DGX Spark gets a 64GB model from $4,999, plus a tool to pair two units

Nvidia said in a blog post on 2 October 2026 that a 64GB version of its DGX Spark desktop AI system will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI "on Friday, Oct. 23, starting at $4,999". The same post announces NVIDIA Sync Cluster Assistant, a feature of the NVIDIA Sync app that links two units together.

What DGX Spark is

On its product page, Nvidia describes DGX Spark as "a complete platform for local autonomous agents", powered by the NVIDIA GB10 Grace Blackwell Superchip and "built to run always-on agent workloads - right from the desktop". The specifications Nvidia lists include:

• Tensor performance: "Up to 1 PFLOP FP4"
• System memory: "64 GB LPDDR5X* or 128 GB LPDDR5x, coherent unified system memory", with the asterisk noting that the 64GB configuration "is available exclusively through participating OEM partners"
• Memory bandwidth: "273 GB/s"
• Storage: "Up to 4 TB NVME.M2 with self-encryption" (the spec table does not say which storage option the $4,999 configuration includes)
• Networking: "ConnectX-7 NIC @ 200 Gbps"
• Power supply: "240 Watts"; dimensions "150 mm L x 150 mm W x 50.5 mm H"; weight "1.2 kg"

The 64GB configuration

The blog says the new configuration "keeps the platform at an accessible price point while retaining the GB10 Grace Blackwell Superchip, DGX OS and full NVIDIA AI software stack — same as the 128GB model". Nvidia says it "supports up to 100-billion-parameter models and the agentic applications built on them, fully on device".

It is sold through partners only: Nvidia's product page says the 64GB model is "available exclusively through participating OEM partners".

The $4,999 figure is Nvidia's US-dollar starting price. Nvidia's post gives no UK price, and we have not converted the dollar figure.

Clustering two units

According to Nvidia, every DGX Spark has a built-in ConnectX-7 network card, and "two units can connect directly with a QSFP cable, pooling their memory to 128GB and expanding model support to up to 200 billion parameters". The company says the pair delivers "twice the memory bandwidth and up to 1.7x the performance". That 1.7x figure comes from what the post calls "NVIDIA's Qwen 3.8 27B test", in which "two clustered 64 GB systems delivered up to 1.7x performance compared with a single system".

The post includes a five-row table image comparing four set-ups (one, two and four units). Its relative-performance row reads 1x for a single 64GB unit, 1.7x for each of two two-unit configurations (128GB and 256GB of memory), and 3.1x for four units with 512GB. The same table lists FP4 compute of 1 PF, 2 PF, 2 PF and 4 PF and bandwidth of 273 GB/s, 546 GB/s, 546 GB/s and 1,092 GB/s for the four set-ups, so the 1.7x and 3.1x figures sit alongside doubled and quadrupled hardware. Those are Nvidia's own figures; Nvidia's post does not cite independent tests, and the 64GB model is not yet on sale.

Nvidia says the cluster assistant "detects connected units, validates device configuration and configures the ConnectX-7 network", and that "nothing needs to be reconfigured when scaling from one unit to two". The post also says a separate tool, NVIDIA Sync Model Launcher, is "coming at the end of the month" to download and launch Qwen3.8 27B on one unit or a cluster.

Who it is for

Nvidia pitches the 64GB model at "developers, researchers and AI enthusiasts". The workflows it gives as examples are:

• running a coding or research agent "around the clock";
• running a language- or image-generation model on the Spark while using a laptop or desktop for other work;
• adding a second unit when, in Nvidia's words, "a single task outgrows one unit".

Nvidia says DGX Spark supports its Agent Toolkit, CUDA-X AI libraries and Nemotron open models, plus runtimes including Ollama, vLLM and PyTorch, "out of the box". It also says Blender is "among the first major creator application providers to support the platform", with an installer "coming soon".

Why it matters

Nvidia pitches the 64GB model on price while keeping the same chip and software. On our reading, the cluster assistant is Nvidia's answer to buyers who outgrow one box. Whether the claimed scaling holds outside Nvidia's own test is something buyers will want to see measured independently once units ship on 23 October.