NVIDIA Adds 64GB DGX Spark for On-device Model Work

Executive Summary

• NVIDIA said a 64GB unified-memory configuration of DGX Spark will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Friday, Oct. 23, according to the company's public statement.
• The new configuration retains the GB10 Grace Blackwell Superchip, DGX OS and the NVIDIA AI software stack, and supports up to 100-billion-parameter models on device, per the same statement.
• Two 64GB units can be clustered through NVIDIA Sync Cluster Assistant to pool 128GB of memory and expand model support to up to 200 billion parameters, with NVIDIA reporting up to 1.7x performance on a Qwen 3.8 27B test versus a single system, the statement said.
• Pricing starts at $4,999, and the configuration is available exclusively from manufacturer partners, the company said.

Key Takeaways

• NVIDIA is widening the DGX Spark line rather than replacing it, adding a lower-memory, lower-priced SKU alongside the existing 128GB model while keeping the same silicon and software base.
• The commercial pitch is local execution of agentic and inference workloads without cloud dependency, with the 64GB unit positioned for models up to 100 billion parameters.
• Scaling is handled through a direct QSFP connection between two units plus the NVIDIA Sync Cluster Assistant, which auto-detects devices, validates configuration and configures the ConnectX-7 network.
• Availability rests on six named manufacturer partners, and NVIDIA frames the 1.7x two-unit figure as its own benchmark result on one model, not a general performance guarantee.

What Changes in the NVIDIA DGX Spark Lineup

The headline fact is a hardware tier, not a new architecture. DGX Spark already combined Grace Blackwell compute, unified memory, a ConnectX-7 network interface and a CUDA-accelerated software stack in a compact system. The 64GB version cuts memory in half relative to the 128GB model and keeps the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA AI software stack. NVIDIA frames the trade-off as accessibility: the new configuration stays at a lower price while preserving the development environment.

That matters because memory capacity is the binding constraint for local model work. NVIDIA states the 64GB unit supports up to 100-billion-parameter models entirely on device, with agentic applications built on those models. Developers, researchers and AI enthusiasts are the named audiences. The system ships with NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and runtimes including Ollama, vLLM and PyTorch with CUDA, which NVIDIA says allows a path from power-on to running models in minutes.

Blender is described as among the first major creator application providers to support the platform, with a prebuilt, downloadable installer coming soon. That is a stated roadmap item, not a shipped capability, and should be read as such.

How NVIDIA Sync Cluster Assistant Handles Scaling

The second element of the announcement is operational rather than silicon-level. Every DGX Spark includes a built-in NVIDIA ConnectX-7 NIC. Two units can be joined directly with a QSFP cable, pooling memory to 128GB, expanding model support to up to 200 billion parameters and delivering twice the memory bandwidth. NVIDIA claims up to 1.7x performance in its Qwen 3.8 27B test when two 64GB systems are clustered, compared with a single system, with room to keep scaling as workloads demand.

The NVIDIA Sync app is what configures that cluster. The cluster assistant feature detects connected units, validates device configuration and configures the ConnectX-7 network. Each node runs the same software stack, so NVIDIA says nothing needs reconfiguring when moving from one unit to two. The company also said NVIDIA Sync Model Launcher is coming at the end of the month, letting developers download and launch Qwen3.8 27B on a single system or a cluster, with the model configured to run across connected devices and accessible from users' laptops. It will also set up OpenCode so developers can start coding in a browser.

The mechanism is worth separating from the marketing. Pooling memory is a capacity gain; the 1.7x figure is a single-vendor benchmark on one model, and NVIDIA does not describe the test conditions, context length or batch configuration. Buyers comparing a two-unit cluster against a single larger-memory machine should treat the multiplier as indicative of direction, not a specification.

Related: Google AI Opens Economy Atlas Data to Public Access in 2026

NVIDIA Workflow Scenarios and Their Limits

NVIDIA lists three practical patterns. The first is an always-on agent that reviews code, analyzes documents or carries out multistep tasks, with a cluster adding capacity for larger models, longer context windows or multiple concurrent agents. The second is a model server on the desk: DGX Spark handles inference while laptops or desktops run a separate agent or creative application. The third is scaling in place, where a task outgrows one unit and two 64GB systems linked over the 200 GbE fabric pool memory without software reconfiguration.

These are described use cases from the vendor, not independently tested deployments. The consistent thread is that the constraint being managed is memory and context, not raw compute alone. For a single developer, 64GB at the entry price covers a defined class of open models. For teams running long-context work or several agents at once, the practical unit of purchase may be a pair rather than one box, which shifts the effective entry cost well above the $4,999 starting figure even though NVIDIA does not publish a cluster bundle price.

NVIDIA Partner Availability and Market Signals

Distribution runs through Acer, ASUS, Dell, Gigabyte, HP and MSI, with availability on Friday, Oct. 23 and an exclusive manufacturer-partner channel. That is a broad set of OEMs for a workstation-class product and indicates NVIDIA is pushing DGX Spark through established PC supply chains rather than a narrow direct channel.

For deeper context, see our Automotive analysis: "How Automotive Is Converging Software, AI and EV Platforms in 2026, According to McKinsey and Gartner".

Reader comments on the announcement, included in the source page, objected to pricing. One commenter called the system overpriced; another argued the price is too high for end users and hoped AMD and Intel would offer competing hardware. These are user opinions captured on the vendor blog, not verified market data, and NVIDIA did not respond to them in the material reviewed. They are nonetheless a signal that the accessible-price framing will be tested against buyer expectations.

The source also notes, as separate "ICYMI" items, that playbooks for vLLM, OpenClaw with a local LLM, and connecting multiple DGX Sparks are coming soon to 64GB devices; that new Windows PCs powered by NVIDIA RTX Spark are coming this month from Acer, ASUS, Dell, HP, Lenovo, Microsoft and MSI; and that Alibaba's Qwen-Image-2.1 runs locally on NVIDIA RTX GPUs, DGX Spark and DGX Station. Those items sit outside the core 64GB announcement and are best treated as adjacent context rather than part of the launch.

What This Means for Practitioners

For developers and research teams weighing local inference against cloud spend, the relevant question is not whether 64GB is enough in the abstract but whether the specific models already in use fit within it. NVIDIA's stated ceiling of 100 billion parameters on a single unit, and 200 billion across two, gives a concrete filter to apply before purchase. Procurement teams should price a two-unit configuration rather than the $4,999 entry point if their workloads need long context or concurrent agents, and should treat the 1.7x cluster figure as a vendor benchmark pending independent testing.

Additional coverage: Latest Clean Tech Market Size and Forecast Statistics 2026-2030

NVIDIA Implementation Risks

The main risks visible in the source material are descriptive rather than speculative. NVIDIA's performance and scaling claims come from its own testing and are stated for one model, Qwen 3.8 27B, without disclosed conditions, so buyers cannot generalize them. Cluster benefits depend on a direct QSFP link between exactly two units and on NVIDIA Sync Cluster Assistant performing detection, validation and network configuration correctly; NVIDIA does not describe failure modes or multi-node behavior beyond two systems in the material provided. The Blender installer and the NVIDIA Sync Model Launcher are described as coming soon or arriving at the end of the month, not as available at launch. Availability depends on six external manufacturer partners meeting the Oct. 23 date. Finally, the source includes no independent verification, customer deployments, return data or third-party benchmarks.

Editorial independence disclosure: this article was prepared from a single vendor-published source and does not reflect independent testing or reporting.

Source note: all facts, figures and product details above are drawn from the NVIDIA Blog post at https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/.

DGX Spark 64GB Signals Table

The source does not specify geography for any entity, and no row has been added beyond what the source explicitly names.