NVIDIA has begun shipping DGX Spark, a compact developer box powered by the GB10 Grace Blackwell Superchip. It delivers ~1 petaFLOP of AI performance with 128GB unified memory and the full NVIDIA AI software stack preinstalled.
📌 Key Takeaways
- Shipping now with a GB10 Grace Blackwell Superchip and 128GB unified memory.
- Runs and fine-tunes models locally, up to 200B parameters; 70B for on-box fine-tunes.
- ConnectX-7 200 Gb/s networking and NVLink-C2C enable low-latency CPU-GPU memory.
- MediaTek co-designed the GB10; two Sparks can team to serve ~405B-param models.
- Systems from Acer, ASUS, Dell, HP, Lenovo, MSI, Gigabyte join NVIDIA’s own unit.
How DGX Spark Runs 200B-Param Models Locally With NVLink-C2C
DGX Spark is positioned as a “personal AI supercomputer” for researchers, engineers, and students who want to run modern reasoning and vision-language models at the desk instead of the data center.
NVIDIA says orders open on its site and through channel partners, with OEM versions landing across major PC brands.
The launch stretches the original “Project DIGITS” concept into production hardware. Spark arrives with the NVIDIA AI stack and NIM microservices so you can pull down models, wire up agents, and ship local workflows out of the box.
“DGX Spark puts an AI computer in the hands of every developer to ignite the next wave of breakthroughs.” — Jensen Huang, CEO, NVIDIA
🎉 To celebrate DGX Spark shipping worldwide starting Wednesday, our CEO Jensen Huang just hand-delivered some of the first units to @elonmusk, chief engineer at @SpaceX 🚀, today in Starbase, Texas.
The exchange was a connection to the new desktop AI supercomputer’s origins –… pic.twitter.com/BZnsl2D2YS
— NVIDIA (@nvidia) October 14, 2025
Specs And Performance
At the heart is the GB10 Grace Blackwell Superchip, coupling an Arm CPU complex with a Blackwell-class GPU through NVLink-C2C, which NVIDIA rates at ~5× PCIe Gen5 bandwidth for coherent memory access.
That design underpins the 128GB unified memory used for larger context windows and multi-modal workloads.
NVIDIA quotes up to ~1 petaFLOP of AI performance (FP4, with sparsity) and says Spark can run models up to 200B parameters locally, while fine-tuning fits up to ~70B. A built-in ConnectX-7 200 Gb/s NIC supports fast data ingest and small clusters.
GB10 Grace Blackwell, 128GB unified memory, NVLink-C2C, 200 Gb/s networking, and ~1 PFLOP AI in a desk-friendly box.
MediaTek’s Role And Scale-Up Options
MediaTek confirms it co-designed GB10 for Spark, focusing on CPU, memory subsystem, and high-speed interfaces for the Grace 20-core Arm CPU. The company also notes a two-node Spark setup can serve ~405B-parameter models using the onboard networking.
This co-development signals a broader ecosystem around die-to-die interconnects and low-power design. It also shows how Spark can be an on-ramp to agentic and physical AI work without immediately renting large clusters.
Software, Models, And Early Access
Spark ships with the NVIDIA AI stack and access to NIM microservices, letting teams stand up local apps like image generation with FLUX.1, vision search with Cosmos Reason, or a Qwen3 chatbot tuned for the box.
These targets reflect the push to bring agents and tools closer to data and teams.
NVIDIA highlights early units in the hands of developers at Anaconda, Docker, Hugging Face, and others to validate tools and model performance on the GB10 platform ahead of wider availability.
How To Get One
Orders begin on NVIDIA.com, with partner systems from Acer, ASUS, Dell Technologies, Gigabyte, HP, Lenovo, and MSI rolling out globally through retail and channel. Check regional timelines, as availability may vary.
If you plan to cluster multiple Sparks or pair with data-center gear, align on 200 Gb/s switching and storage throughput early. That avoids bottlenecks when you move from single-box prototyping to multi-node inference.
Why It Matters
Local fine-tuning and evaluation shorten iteration loops and reduce cloud spend during early model work. For teams dealing with sensitive data, on-desk workloads can also simplify privacy and residency concerns while you validate ideas.
As more OEMs adopt GB10, expect a wider menu of form factors. The useful test will be sustained throughput under real agents and tool-use, not just headline TOPS. Spark’s unified memory and stack integration are the practical bets to watch.
Conclusion
DGX Spark brings petaflop-class compute and a curated software stack to the desktop. If the ecosystem holds and small clusters remain simple to deploy, this form factor becomes a common path to build and ship agentic AI.
The next signals are OEM ship dates, two-node performance on >200B models, and how quickly teams push from prototype to production without changing code paths.
For the recent AI News, visit our site.
If you liked this article, be sure to follow us on X/Twitter and also LinkedIn for more exclusive content.
