#

Key takeaways

• NVIDIA announced a 64GB unified-memory version of DGX Spark.

• The system keeps the GB10 Grace Blackwell Superchip and NVIDIA AI software stack.

• Two 64GB systems can be clustered to pool memory to 128GB.

• NVIDIA says the setup can support models up to 200 billion parameters in a cluster.

• Availability begins October 23 from named manufacturer partners.

NVIDIA adds a 64GB DGX Spark option

NVIDIA announced a 64GB unified-memory version of DGX Spark on October 2, 2026. The company says the new configuration will be available beginning October 23 from Acer, ASUS, Dell, Gigabyte, HP, and MSI. NVIDIA lists the starting price at $4,999.

The new model keeps the GB10 Grace Blackwell Superchip, DGX OS, and NVIDIA AI software stack used by the 128GB version. NVIDIA presents the machine as a compact local platform for AI agents, inference, fine-tuning, data science, and edge development. The source does not add more detail about hardware changes beyond the memory configuration.

What NVIDIA says the 64GB model can do

According to NVIDIA, the 64GB system can support models with up to 100 billion parameters fully on device. The company adds an important condition: this depends on the model and workload requirements. That means the claim is tied to NVIDIA's own specifications, not a universal result for every use case.

NVIDIA also says the system works with Ollama, vLLM, PyTorch with CUDA, its Agent Toolkit, and Nemotron models. A model launcher is planned for later in the month. The source does not provide more detail on how that launcher will work.

Clustered memory and scaling

NVIDIA Sync Cluster Assistant can join two 64GB systems over their ConnectX-7 network interfaces. NVIDIA says this pools memory to 128GB, increases available memory bandwidth, and expands supported model size to as much as 200 billion parameters. The company describes this as a way to scale local AI work across two systems.

In a company test using Qwen 3.8 27B, two clustered machines reached up to 1.7 times the performance of one. This is a measured claim from NVIDIA, not an independent benchmark across all workloads. The result should be read as a product-specific test rather than a general guarantee.

The assistant also detects attached systems, validates configuration, and configures networking. That suggests NVIDIA is trying to reduce setup friction for users who want to link two devices. The source does not say whether the process requires manual steps beyond those functions.

Operational considerations

The main operational point in the source is that performance and model support depend on workload requirements. Users should treat the published limits as product guidance, not a promise for every model or task. The clustering feature also depends on compatible networking and the assistant's configuration process.

The source is clear that these are NVIDIA's product specifications and measured claims. It does not establish independent performance across all workloads. It also does not say anything about broader deployment conditions outside the named product details.

Morocco relevance

The source reports no Morocco-specific availability or local market detail. For readers, the global lesson is simple: check whether a product claim is tied to a specific configuration, workload, or test setup before planning adoption.

Bottom line

NVIDIA's 64GB DGX Spark adds a lower-memory entry point to the same local AI platform. It keeps the core software stack, supports a wide set of AI tools, and can scale through two-system clustering. The source frames the product as a compact option for local AI work, with performance claims that depend on NVIDIA's own test conditions.