NVIDIA is adding a 64GB model to its desktop AI system DGX Spark, going on sale on October 23. It starts at 4,999 USD (about 790,000 yen) and will be sold through six partners: Acer, ASUS, Dell, GIGABYTE, HP and MSI. A single unit can run models of up to roughly 100 billion parameters, and two units linked by cable behave as a 128GB memory pool.

Where the 64GB model fits

DGX Spark is a compact AI development machine built around NVIDIA's GB10 Grace Blackwell Superchip. The new model carries 64GB of unified LPDDR5X memory with 273GB/s of bandwidth, plus a ConnectX-7 network interface rated for up to 200Gbps.

The price stands out. The 128GB model first arrived at 3,999 USD, yet the new 64GB version, with half the memory, starts at 4,999 USD. Reports point to rising memory costs as the backdrop, while NVIDIA says it is aiming to keep the platform at an accessible price point. ※1 USD = 158 JPY (approximate, based on the exchange rate level on October 5, 2026)

Up to 100B-class models on a single unit

According to NVIDIA, the 64GB version still handles models of up to about 100 billion parameters. Popular models in use today, such as Qwen 3.8 27B and Nemotron 3.5 Lightning, fit within that range. The impact is most likely to show up with models around 80 billion parameters or larger, and with fine-tuning work.

For running LLMs locally or testing agents, the halved memory should matter relatively little.

Two units, 128GB

The other key point is clustering. Two 64GB units connected directly with a QSFP cable, with no switch required, pool their memory into 128GB and support models of up to about 200 billion parameters. In NVIDIA's own Qwen 3.8 27B test, the clustered pair delivered up to 1.7 times the performance of a single 128GB system. Note that this figure comes from NVIDIA's own measurements.

Up to four units can be linked at 200Gbps each. NVIDIA Sync Cluster Assistant automates the setup, and Model Launcher, which deploys models across the cluster, is due at the end of October.

Software and availability

The system runs DGX OS and gives access to CUDA-X AI libraries and the NVIDIA Agent Toolkit. It also supports standard local inference tools such as llama.cpp, Ollama, vLLM and LM Studio.

Sales go mainly through the six partners above rather than directly through NVIDIA. Pricing and availability in Japan have not been confirmed so far.

Summary

The DGX Spark 64GB offers a lower entry point through reduced memory while keeping room to grow through two-unit clustering to 128GB. The price rose rather than held steady, but developers who want to try 100B-class models locally gain another option. Japanese availability remains to be announced.