Nvidia is adding a 64GB version of its DGX Spark desktop AI computer, priced from $4,999 and due to go on sale Oct. 23, the company said Friday in a blog post. Two of the new machines can be linked with a single cable to share 128GB of memory, a setup Nvidia says can handle models of about 200 billion parameters.

What Happened

The 64GB model uses the same GB10 Grace Blackwell superchip, built-in ConnectX-7 networking and DGX OS as the existing 128GB DGX Spark, and it ships with the same Nvidia AI software stack preinstalled. Nvidia said a single unit supports models of up to roughly 100 billion parameters.

The machine will be sold through Acer, ASUS, Dell, Gigabyte, HP and MSI. Nvidia gave $4,999 as the starting price.

RelatedClaude Code Desktop Preview Flaw Could Have Exposed Developers’ Cloud Keys, Researchers Say

Clustering is handled by a tool called Nvidia Sync Cluster Assistant, which configures two 64GB units connected by a QSFP cable. Pooled, the pair offers 128GB of memory and twice the memory bandwidth of one machine, Nvidia said. In the company's own tests with the Qwen 3.8 27B model, two clustered systems delivered up to 1.7 times the performance of one.

Why It Matters

Memory is the main constraint on which models a desktop machine can run locally. By selling a half-memory version and letting buyers add a second unit later, Nvidia lowers the upfront cost for developers who want to try open models and AI agents on their own hardware rather than on rented cloud GPUs. A team that outgrows one box can pair two and reach the same 128GB capacity as the larger model, with double the bandwidth.

The software bundle is pitched at agent development. Nvidia lists its Agent Toolkit, CUDA-X libraries and Nemotron models, along with support for the open-source runtimes Ollama, vLLM and PyTorch with CUDA. For companies that do not want source code or internal data leaving the building, a local machine that runs these tools out of the box is an alternative to sending prompts to a hosted model.

Coming Software

Nvidia said a separate tool, Sync Model Launcher, will arrive at the end of October. It is designed to download and run Qwen 3.8 27B on either one unit or a two-unit cluster, and to set up OpenCode, an open-source coding agent, for use in a web browser. That would give buyers a working local coding assistant without manual installation of models and serving software.

The company is also courting creative users. Blender is among the first creative applications to support the platform, and Nvidia said a prebuilt, downloadable installer is coming soon.

What to Watch

Partner listings and pricing should become clear when the systems go on sale Oct. 23. Developers will look for independent benchmarks of the clustered setup, since the 1.7 times figure comes from Nvidia's tests on one model and real gains will depend on model size and workload. The end-of-month release of Sync Model Launcher will show how much of the setup Nvidia has actually automated, and whether support extends beyond the single Qwen model named in the announcement.