On October 2, 2026, NVIDIA unveiled the NVIDIA DGX Spark 64GB, the latest member of NVIDIA’s lineup of compact AI desktops. Available from $4,999, the new configuration continues to feature NVIDIA’s GB10 Grace Blackwell SuperChip and is aimed at developers looking to deploy large AI models locally. NVIDIA claims the system can train models with up to 100 billion parameters on just one system, with a pair of 64GB systems joined together to form a 128GB memory pool.
NVIDIA DGX Spark 64GB Specs: What You Get
The DGX Spark 64GB uses the same core platform as the larger configuration rather than introducing a slower processor or GPU.
Key specifications include:
- GB10 Grace Blackwell Superchip
- 64GB unified LPDDR5X memory
- Blackwell GPU architecture
- 20-core Arm CPU
- 5th-generation Tensor Cores
- Up to 1 PFLOP FP4 AI performance with sparsity
- NVIDIA DGX OS
- NVIDIA CUDA-X AI software stack
- ConnectX-7 networking
- Local AI inference, development and agent workloads
NVIDIA’s documented DGX Spark platform reveals 273GB/s of memory bandwidth; the new 64GB version has the same GB10-based architecture.
The system is compact to fit in any workplace — with dimensions of 150 × 150 × 50.5mm and weighing approximately 1.2kg — a fraction of the size of conventional AI workstations.
NVIDIA DGX Spark 64GB Price and Release Date
The NVIDIA DGX Spark 64GB price starts at $4,999.
The new configuration will only be offered by hardware partners such as Acer, ASUS, Dell, Gigabyte, HP and MSI, unlike NVIDIA’s Founders Edition offering. It will be available on October 23, 2026.
It’s therefore a new entry level configuration of the 64GB over the 128GB system.
Can DGX Spark 64GB Run 100B AI Models?
NVIDIA’s 100B-Parameter Claim
Yes, replies NVIDIA. The 64GB DGX Spark is capable of running local AI models with up to 100 billion parameters, the company claims. Note that this is by no means a guarantee that every 100B model will achieve a high speed or have unrestricted context.
The usable capacity and performance depends on the model size, quantization, KV-cache and inference software.
NVIDIA notes that models with up to 200 billion parameters can run on the 128GB DGX Spark for inference.
DGX Spark 64GB Benchmarks: What We Know
There is an important distinction between model capacity and benchmark performance.
NVIDIA has not yet announced a universal benchmark for its newly announced 64GB system, but it has released a 100B model-tokens-per-second benchmark already. Instead, it is Qwen 3.8 27B that heads the list of performance results.
NVIDIA says two 64GB DGX Spark systems cluster up to 1.7× the performance of one, on its Qwen 3.8 27B test.
There are about 25–40 tokens per second of independent community testing results on DGX Spark hardware for Qwen3.8 27B based on quantization, inference engine, context length and speculative-decoding setup. These are basically taken on systems that are 128GB, so don’t be misled into thinking that they are 64GB benchmarks.
Whereas early 100B benchmark claims should be taken with a grain of salt until the 64GB retail version is available for repeat testing.
Quick updates on tech, AI, and digital trends.
Why Did Philippines Lift Discord Ban?
Check out Discord’s new child protection commitments.
Can Gemini Really Guess Passwords?
Find out how Gemini breached companies in testing.
What Makes One UI 9 Stand Out?
Uncover Galaxy AI, scam detection, and camera upgrades.
Could Anthropic Reach $2 Trillion?
Browse its IPO plans, Nvidia investment, and revenue surge.
What Are iPhone Prices In Philippines?
See storage options, official prices, and reseller discounts.
Two DGX Spark 64GB Systems Can Scale Up
One of the most interesting features is NVIDIA’s Sync Cluster Assistant.
The 64GB systems can be paired together using the networking hardware for a pooled memory configuration of 128GB. This can be used to run models with up to 200 billion parameters, and serves as a way for developers who have workloads that exceed the capacity of a single machine.
This proves to be a more than just a desktop AI computer with less memory. It’s also a gateway to the local AI scaling ecosystem provided by NVIDIA.
FAQs
What is the NVIDIA DGX Spark 64GB price?
The announced starting price is $4,999, with availability from OEM partners beginning October 23, 2026.
Can NVIDIA DGX Spark 64GB run a 100B model?
NVIDIA says it can run up to 100-billion-parameter models locally. Actual performance depends on quantization, context size and the specific model.
Is there an official 100B benchmark for DGX Spark 64GB?
Not yet. NVIDIA has stated the 100B model-capacity figure, but its published performance example for the new configuration focuses on Qwen 3.8 27B and two-system scaling.
When does DGX Spark 64GB launch?
The new configuration is scheduled to become available from participating manufacturers on October 23, 2026.
