The three tiers, in plain terms
| Tier | OS | Intended user | Headline claim |
|---|---|---|---|
| RTX Spark | Windows | Individuals running personal agents, creators, mobile gamers | Up to 128GB unified memory, 1 petaflop AI compute |
| DGX Spark | Linux | Developers building and scaling apps | Clusters of up to eight units, 1TB unified memory |
| DGX Station | Linux now, Windows in December | Enterprise teams | 748GB coherent memory, up to 20 petaflops |
RTX Spark: agents on the Windows PC
NVIDIA's blog describes RTX Spark as a platform for personal agents, content creation and 1440p gaming at 100 frames per second. It lists up to 128GB of unified memory and 1 petaflop of AI compute. The company dates its release to October 16.
Tech Critter says RTX Spark is shipping "across the world". NVIDIA's post only supports a narrower claim. It says laptops from ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI are available today, with Acer coming soon. Treat availability as vendor- and region-dependent until you see stock where you live.
Two details matter to Windows enthusiasts:
- Underdog is a new desktop agent app, available now. NVIDIA says its Husky inference engine runs models entirely on the device, so user data never leaves the PC. That is the vendor's description of how the app works. It is not an independent privacy audit. If privacy is why you want a local agent, check the app's network behavior yourself.
- Unlocked bootloaders. NVIDIA says many RTX Spark laptops ship with an unlocked bootloader so developers and enthusiasts can customize their local agents. "Many" is not "all". Check the specific model before you buy if this matters to you.
DGX Spark: 64GB models and eight-node clusters
New 64GB DGX Spark systems begin rolling out from partners on October 23. NVIDIA says they can run models including Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Guru3D adds a caveat I agree with. What fits depends on quantization, runtime and context size. A model appearing on NVIDIA's list does not mean it fits comfortably in every configuration.
The bigger story is clustering. DGX Spark now supports clusters of up to eight units in manual configuration. That covers larger models, longer contexts, concurrent agents and distributed development. Wccftech notes this is a big step up from the four-way clusters previously supported. The units connect through built-in ConnectX-7 networking and the NVIDIA Sync Cluster Assistant.
The word "manual" matters. This is not a turnkey appliance. NVIDIA's technical blog says eight-node support in NVIDIA Sync is coming by the end of the year. That is later than the "this month" timing for the Model Launcher and eight-unit support in NVIDIA's announcement post. Anyone planning a build should expect to wire and configure the cluster by hand for now.
What NVIDIA's own benchmarks show
NVIDIA's technical blog from October 22 gives the memory math. A single DGX Spark has 128GB, four nodes reach 512GB, and eight reach 1TB. NVIDIA says the full GLM 5.3 checkpoint, 753 billion parameters and about 470GB in NVFP4, runs across the eight-node cluster. In its test, GLM 5.3 Flash produced roughly 80 tokens per second.
For the full GLM 5.3 model, NVIDIA reports decode throughput of 31 tokens per second at concurrency one. That rises to 67 tokens per second in total at concurrency eight. Per-user speed drops to 18 tokens per second at that load. The decode scaling test used Qwen3.8-27B in NVFP4 with speculative decoding off and concurrency one:
- One node: 11.86 tokens per second.
- Two nodes: 21.71.
- Four nodes in a physical ring: 32.82.
- Four nodes through a switch with the RoCEnante backend: 36.34.
- Eight nodes in a ring: 40.11.
- Eight nodes through a switch with RoCEnante: 52.58, which is 4.43 times the single-node result.
The scaling is real but sublinear. Eight machines give roughly four and a half times the decode speed of one. NVIDIA also reports a nanochat pretraining run that reached 7.79 times the throughput on eight nodes, at 97.4% scaling efficiency. Training scales better here because nodes mainly synchronize gradients once per step.
Two practical points from the same post:
- Ring or switch? A switchless ring is cheaper and simpler. NVIDIA recommends it for a four-node inference box. A switch is needed to scale to eight nodes, to split the cluster into groups, or to run any-to-any traffic patterns.
- RoCEnante is not a universal fix. It applies only to switch-connected topologies. NVIDIA says NCCL stays the default and is the better choice as cluster size and concurrency grow. The backend falls back to NCCL if its startup test fails.
Software coming this month
- NVIDIA Sync Model Launcher is due later in October. NVIDIA says it will download a model and container suited to the configuration and connect it to OpenCode. It will also load DGX Spark Skills for troubleshooting, maintenance and documentation access.
- The October DGX Spark update adds a Wi-Fi Assistant for switching networks while the system runs in appliance mode. It also refreshes the Skills.
DGX Spark suppliers named by NVIDIA are Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI and NVIDIA.
DGX Station: manageability and a Windows edition
DGX Station offers 748GB of coherent memory and up to 20 petaflops. NVIDIA says that is enough to run models like GLM 5.3 and serve 30+ users at cloud-comparable speeds. Treat the cloud-comparable claim as marketing until someone tests it. NVIDIA Sync now supports DGX Station and links two Linux-based systems.
Intune and six other management integrations
This is the part IT administrators should note. NVIDIA announced seven new integrations: Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Chef and Tanium. NVIDIA says they let IT teams fit DGX Station into enterprise setup, management and compliance processes. It also says they help speed approvals and reduce deployment risk.
NVIDIA's announcement does not say which Intune policies are supported, how enrollment works, or which operating system editions are covered. Until documentation appears, "Intune integration" tells you the direction of travel. It does not tell you what you can manage.
DGX Station for Windows, expected in December
NVIDIA says Windows-based systems will let engineers run local models and agents alongside Windows-native professional software. NVIDIA describes the hardware as running on RTX PRO GPUs, with LLM agents on GB300 systems, all in one machine. Systems are expected from ASUS, Dell Technologies, GIGABYTE, HP, MSI and Supermicro.
Announced application support covers nTop, Dassault Systèmes, Keysight Technologies, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera. Agent options named are Hermes Agent, Perplexity and Aible, running on premises.
The pitch is that much engineering software only runs on Windows, and today a local-AI box has often meant a separate Linux machine. Whether that holds up depends on pricing, driver and agent stability, and how well the Windows image works with endpoint management. Those details are not in the announcement.
Existing Linux DGX Station systems are shipping now from ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro and Exxact.
Other Station items
- Three open-source robotics reference workflows bring GPT-6 Astra, Isaac Lab, Isaac Lab-Arena and Cosmos to DGX Station. They cover refining GR00T 1.7 models, training dexterous manipulation skills and post-training Cosmos world models.
- NVIDIA Cloud Functions comes to early access on Linux DGX Station. Teams register local model endpoints in a shared catalog behind one OpenAI-compatible entry point. NVIDIA says developers can swap the model behind an endpoint without changing application code. Teammates can use shared models without direct machine access.
- A DSX OS preview and a local token-factory demo for more than 30 developers will be on the GTC Berlin show floor.
Models and inference
- Recursive Flint is an open-weights coding model from Recursive Superintelligence under Apache 2.0. It has 40 billion total parameters and 2.5 billion active. NVIDIA says it runs on RTX Spark, DGX Spark, GeForce RTX and RTX PRO GPUs. It attributes a 55% generation-speed gain on RTX Spark to custom kernels from Recursive's automated research system, with the weights unchanged.
- Day-zero support covers Nemotron Bolt E4B (as little as 4GB of memory, releasing October 23), GLM, DeepSeek, Qwen, MiniMax M3.1 Flash, Liquid AI's LFM2.5-24B-A2B and FLUX.3 Klein. Klein is a 5.6-billion-parameter image model. NVIDIA says it runs on GeForce RTX GPUs with 8GB of VRAM and makes a 1-megapixel image in about four seconds on an RTX 5090.
- Optimized checkpoints are quantized to save RAM and leave room for KV cache. The first are Qwen3.8-Flash-Next-FP4-Local, for a single 128GB DGX Spark or RTX PRO 6000, and GLM-5.3-Flash-FP4-Local, for two systems.
- Inference engines: llama.cpp speculative decoding with MTP is up to 1.4 times faster for Qwen 3.8 Flash Next agentic tasks on RTX PRO 6000 Blackwell. vLLM is up to 1.3 times faster for Qwen3.8 27B on DGX Spark. GLM 5.3 Flash gets up to 1.2 times higher throughput on DGX Station.
Every "up to" figure is NVIDIA's own and depends on workload. Some outlets restate these as flat percentage gains. That drops the "up to", which is the part that matters.
What to take from this
- Windows PC buyers: RTX Spark laptops are on sale from named vendors, and Underdog is available now. Verify the privacy claim and the bootloader status for your exact model.
- Developers: Eight-node clusters raise the memory ceiling to 1TB, but you build them manually and the scaling is sublinear for inference. Choose a ring for four nodes. Choose a switch for eight nodes or split workloads.
- IT admins: Intune appears on the DGX Station integration list, but there is no policy or deployment detail yet. Watch for documentation before December's Windows edition.
- Everyone: NVIDIA's blog says RTX Spark "launched last week", and Tech Critter relays the vendor view. Independent testing of the cluster and Station claims is still to come.
References
- NVIDIA @ GTC Berlin 2026: Big upgrades to DGX Spark + Station and RTX Spark - Tech Critter Tech Critter · 2026-10-23T05:37:52+00:00
- NVIDIA Gives DGX And RTX Systems A 40% AI Inference Boost, While DGX Spark Now Clusters Up To 8 Units For 1 TB Memory wccftech.com
- Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents blogs.nvidia.com