The Pulse
NVIDIA Adds $4,999 64GB DGX Spark for Local Agents
NVIDIA announced a 64GB version of its DGX Spark desktop AI system, priced from $4,999 and scheduled to reach six PC makers on October 23, 2026. The company says two units can pool memory for larger models and agent workloads.
AI.info Team · Written with AI ·
A smaller-memory model with a bigger-workload pitch
NVIDIA’s new DGX Spark configuration cuts the system’s unified memory from 128GB to 64GB, but the company is pitching it as a starting point for local AI workloads that can grow across multiple machines. Announced October 2, the version will start at $4,999 and go on sale October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI.
The lower-memory model retains the GB10 Grace Blackwell Superchip, DGX OS and NVIDIA’s AI software stack. NVIDIA says a single 64GB unit can run models with up to 100 billion parameters entirely on-device. The company’s pitch is that developers can add a second machine when one unit no longer has enough memory for a model, a longer context window or concurrent agent requests.
“The new SKU runs capable local agents on device — privately, without cloud dependency.”
Allen Bourgoyne, NVIDIA Blog
Two units pool memory over a 200 GbE link
Two 64GB Sparks can connect through their built-in ConnectX-7 networking and a QSFP cable, with NVIDIA Sync Cluster Assistant configuring the connection. NVIDIA says the pair pools 128GB of memory and supports models with up to 200 billion parameters. The company describes the system as delivering twice the memory bandwidth of one unit.
NVIDIA also reports a performance result from a specific test: two clustered 64GB systems ran Qwen 3.8 27B at up to 1.7 times the performance of a single system. The announcement does not give further benchmark conditions in the post, so the figure should be read as NVIDIA’s result for that test, not as a general speedup for every model or workload.
Sync handles setup; the hardware still sets the limits
The Sync Cluster Assistant detects connected units, checks their configuration and sets up the ConnectX-7 network. NVIDIA says both systems run the same software stack, so developers do not need to reconfigure the environment when expanding from one Spark to two. The stated use cases include longer-context work, larger models and multiple agents running at once.
For developers who want to start with one system, NVIDIA lists support for the Agent Toolkit, CUDA-X AI libraries, Nemotron models, Ollama, vLLM and PyTorch with CUDA. It also names llama.cpp and LM Studio among supported inference frameworks. These software options do not change the product’s memory ceiling: workloads that exceed what 64GB can accommodate need a second unit or a different system.
Availability begins October 23
The 64GB configuration will be sold exclusively through manufacturer partners, rather than as NVIDIA’s existing 128GB configuration. NVIDIA says the starting price is $4,999 and the six listed partners expect to offer it from Friday, October 23. The company also says its Sync Model Launcher is due at the end of October, with support for launching Qwen3.8 27B on one Spark or across a cluster.
The announcement sets out a clear tradeoff: a lower-cost entry point with half the memory of the 128GB model, plus a vendor-managed route to 128GB when two units are connected. Buyers will need to weigh the $4,999 starting price against the workload they intend to run—and the cost of adding a second system if a single 64GB device proves too small.