NVIDIA adds a 64GB DGX Spark at $4,999, shipping October 23, with two units clustering to 128GB

Clustering two units is the headline pitch of this launch. NVIDIA states that every DGX Spark ships with an onboard ConnectX-7 NIC, so a pair can be connected directly with a QSFP cable over a 200GbE link with no switch in between, while the Cluster Assistant feature in NVIDIA Sync discovers the units, validates device configuration and sets up the ConnectX-7 network. NVIDIA's internal test number uses Qwen 3.8 27B: two clustered 64GB systems delivered up to 1.7x the performance of a single system. NVIDIA also notes that because every node runs the same CUDA libraries, inference engines and DGX OS image, scaling does not require rebuilding the software environment — the only step from one unit to two is the cable. Supported inference frameworks listed are llama.cpp, Ollama, vLLM and LM Studio.

The software half of the announcement has its own end-of-month date. NVIDIA says the NVIDIA Sync Model Launcher arrives at the end of October, able to download and launch Qwen3.8 27B on a single Spark or a clustered pair, with NVIDIA Sync configuring the model across connected devices and exposing the endpoint to the developer's own laptop. It will also set up OpenCode against that model so developers can write code in a browser while the Spark does the inference. DGX Spark ships with the NVIDIA Agent Toolkit, CUDA-X AI libraries and Nemotron open models, and Blender is among the first creator applications to support the platform, with a prebuilt downloadable installer described as coming soon. NVIDIA's three example workflows are worth copying: keep a coding or research agent running around the clock to review code and analyze documents; run the model on the Spark while agents and creative apps run on your own laptop, moving inference off the PC; and add a second unit when a single task outgrows one machine, scaling the same workflow without reconfiguring the environment.

Pricing is where this announcement is easiest to misread, so keep the sources separate. NVIDIA's own line is one sentence: starting at $4,999, sold exclusively through manufacturing partners, who set their own configurations and regional pricing. VideoCardz adds the price history: the 128GB Founders Edition carried an original $3,999 MSRP, NVIDIA raised that to $4,699 in February 2026 citing worldwide memory supply constraints, and the 128GB model's market price has now crossed $6,000. On those numbers, the new 64GB entry point of $4,999 lists $300 above NVIDIA's own current 128GB MSRP. StorageReview puts it more bluntly: the pitch for the cheaper box is that it is cheap relative to market rates on the larger model. One caveat: partner systems and the first-party Founders Edition are not directly equivalent products, since partners determine their own configurations and regional pricing.

For anyone running local inference, three boundaries are worth holding onto. First, the 64GB and 128GB models share the GB10 Grace Blackwell chip and the full software stack; the difference is memory and the model ceiling that comes with it — 100 billion parameters on one unit, 200 billion across two. Second, clustering is the officially promoted path, but it requires each unit's onboard ConnectX-7 plus a QSFP cable, and the 1.7x figure comes from NVIDIA's own internal testing rather than an independent reproduction. Third, the Model Launcher due at the end of October is what turns downloading, cross-device configuration and endpoint exposure into a few clicks; until then you still install your own inference framework and pull your own weights. NVIDIA also notes that playbook pages are live on build.nvidia.com covering four agent options: NemoClaw, OpenClaw, Hermes Agent and OpenShell.