TL;DR: At GTC Berlin, NVIDIA said DGX Spark clustering now supports up to eight units in manual configuration, giving 1TB of unified memory, enough, it says, to run frontier-class open models like GLM 5.3. NVIDIA Sync will add eight-system support and a one-click Model Launcher later this month, and the cheaper DGX Spark 64GB rolls out from partners on Friday, October 23. Performance claims so far are NVIDIA’s own and tied to specific models.
NVIDIA’s desktop AI box just got a lot more room to grow. In a GTC Berlin blog post published October 21, the company detailed DGX Spark clustering of up to eight systems, new software tools and a raft of updates across its local AI lineup. Here is what changed, what it means for developers and what is still missing.
DGX Spark clustering: from two units to eight
DGX Spark systems have built-in NVIDIA ConnectX-7 networking, and developers could already link several of them using the NVIDIA Sync Cluster Assistant. NVIDIA’s Sync documentation says the Cluster Assistant currently supports two to a maximum of four devices.
The new step is eight. NVIDIA says DGX Spark “now supports clusters of up to eight units in manual configuration, enabling larger models, longer contexts, concurrent agents and distributed development.” With eight 128GB units, that adds up to 1TB of unified memory. Notebookcheck notes that a cluster of the new 64GB models would top out at 512GB.
Who NVIDIA says it is for
The company pitches the eight-unit setup as “ideal for small teams sharing a single cluster as a backend, with each developer running concurrent sessions against the same model.” In other words, a shared, private inference server that sits in the office rather than a cloud account. That fits a wider shift toward local AI that we saw with Perplexity Computer running fully on a Mac.
| Configuration | Unified memory | Setup |
|---|---|---|
| 1 x DGX Spark 64GB | 64GB | Single unit |
| 1 x DGX Spark 128GB | 128GB | Single unit |
| 2-4 x DGX Spark | Up to 512GB (128GB units) | NVIDIA Sync Cluster Assistant |
| 8 x DGX Spark 64GB | 512GB | Manual for now |
| 8 x DGX Spark 128GB | 1TB | Manual for now |
Memory totals based on NVIDIA’s and Notebookcheck’s figures.
NVIDIA Sync Model Launcher is coming this month
Manual setup is the catch today, and NVIDIA says software will close that gap. According to the blog, NVIDIA Sync will support eight DGX Sparks and add a Model Launcher “this month.” NVIDIA describes Model Launcher as a tool that, “with a few clicks,” downloads a model and container suited to your configuration, whether a single Spark or a cluster, and connects the model to the OpenCode coding tool. It also loads DGX Spark Skills for troubleshooting, device maintenance and documentation.
The earlier DGX Spark 64GB announcement previewed the same idea, saying the launcher would set up OpenCode so developers can start coding in their browser.
DGX Spark 64GB arrives October 23
NVIDIA says new DGX Spark 64GB systems start rolling out from partners on Friday, October 23. The company lists models it can run, including Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B. Notebookcheck reports the 64GB version keeps the GB10 Grace Blackwell Superchip and starts at $4,999. DGX Spark systems are sold by Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI and NVIDIA.
Other DGX Spark and local AI updates
The Berlin announcements went well beyond clustering. Highlights from NVIDIA’s list:
- October software update: a Wi-Fi Assistant for switching networks in appliance mode, plus updated DGX Spark Skills inside NVIDIA Sync via OpenCode.
- Faster inference: NVIDIA says new vLLM speculative decoding and FlashInfer prefill kernels make agentic tasks with Qwen3.8 27B up to 1.3x faster on DGX Spark, and llama.cpp speculative decoding speeds Qwen 3.8 Flash Next tasks up to 1.4x on RTX PRO 6000 Blackwell.
- Memory-optimized checkpoints: quantized models such as Qwen3.8-Flash-Next-FP4-Local, designed to leave more memory for the KV cache. NVIDIA says it targets a single DGX Spark 128GB, while GLM-5.3-Flash-FP4-Local runs on two systems.
- Day-zero model support: new Nemotron, GLM, DeepSeek, Qwen, MiniMax and FLUX.3 models.
- Recursive Flint: a 40-billion-parameter mixture-of-experts coding model with 2.5 billion active parameters, released by Recursive Superintelligence with open weights under Apache 2.0.
DGX Station and RTX Spark news
NVIDIA also said the larger DGX Station, with 748GB of coherent memory, gets seven new IT management integrations, and that a Windows version arrives from partners in December. On the consumer side, the company highlighted the Underdog desktop agent for RTX Spark PCs. We covered the RTX Spark launch in RTX Spark laptops: prices, models and first reviews.
What’s still unproven
Treat the numbers carefully. The speed-ups NVIDIA quotes apply to specific models and workloads, not across the board. As Notebookcheck points out, NVIDIA “has not provided corresponding eight-system scaling benchmarks,” so it is unclear how efficiently a full eight-unit cluster handles demanding jobs. Eight units at $4,999 or more each is also a serious outlay, before networking and setup time.
Questions buyers should ask
- How well does throughput scale from four to eight units on the model you actually use?
- How stable is manual configuration until Sync support arrives?
- Would a single larger workstation, or cloud GPUs, be cheaper for your team’s usage?
Why DGX Spark clustering matters
The appeal is control. A team that wants to run large open-weight models privately, keep data off third-party servers and avoid per-token bills can now scale DGX Spark to the memory range where frontier-class models fit. If Model Launcher delivers the one-click setup NVIDIA promises, the barrier drops further. If you are new to the jargon, our explainers on context windows and the open-source DeepSeek DeepInfer engine explain why memory and inference software matter so much for running models locally.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
