NVIDIA said RTX Spark compact desktops from ASUS, Dell and MSI are on sale November 18, 2026, turning months of Spark laptop news into desk hardware you can buy today. The announcement pitches a dedicated Windows PC for local agents, image and video generation, and fine-tuning while a laptop handles travel — with up to 128GB unified memory and NVIDIA's AI software stack pre-wired for Hermes, ComfyUI, and Adobe native Windows on Arm builds.
If you followed RTX Spark launch software and Llama for Windows posts, this is the availability milestone: OEM mini PCs, not only spec slides.
TL;DR: what ships November 18
| Item | NVIDIA statement (Nov 18, 2026) |
|---|---|
| Available now | Compact desktops from ASUS, Dell, MSI |
| Coming soon | Acer, GIGABYTE, HP mini PCs; Surface RTX Spark Dev Box |
| CPU / GPU | 20-core Grace CPU, Blackwell RTX, 6,144 CUDA cores |
| Memory | Up to 128GB unified |
| Networking | 10GbE for large model and asset transfers |
| Also announced | Blender 5.3 DLSS viewport, Adobe TensorRT-RTX updates, Ignite Windows ML items |
Why a desktop Spark exists
NVIDIA frames always-on agents and scheduled renders as workloads that should not compete with battery life on a laptop. A compact desktop can keep Hermes or NemoClaw tasks running, continue ComfyUI video pipelines, and host Unsloth fine-tunes while you close the lid elsewhere.
The blog repeats hardware claims from earlier Spark messaging:
- Run 120B-parameter class models locally (quantization and stack dependent), as the NVIDIA RTX Spark product page also states.
- Real-time work on 90GB+ 3D scenes and 4:2:2 video editing.
Those are vendor performance narratives; treat them as targets until you benchmark your models. Our PAIR and Hermes desktop guide covers one-click local models patterns that pair with this hardware wave.
Software stack bundled with the hardware story
NVIDIA ties RTX Spark desktops to an expanding Windows creative and dev ecosystem:
- Hermes Agent and NemoClaw blueprints for local agents.
- Switchyard in Hermes to route tasks between local and cloud models, managing API spend on creative workflows.
- Unsloth for fine-tuning supported models.
- ComfyUI features such as RTX Video Frame Generation and RTX Video HDR, extending generated clips to 60 fps and HDR inside one workflow.
The post describes an Adobe MAX demo chain: Photoshop and ComfyUI through Hermes, sketch-to-render and on-device animation examples. Adobe sections list TensorRT-RTX acceleration in Premiere, Photoshop Remove Tool, Camera Raw on Arm, Substance 3D Painter 12.2 public beta on RTX Spark, and Illustrator / InDesign native Windows on Arm builds with claimed up to 30% gains versus emulated installs — all Adobe and NVIDIA statements, not independent benchmarks.
Blender and creator tooling
Blender 5.3 adds DLSS 4.5 Ray Reconstruction to the Cycles viewport, which NVIDIA markets as near-final-quality lighting while navigating scenes. Requirements: Blender 5.3 plus current Game Ready or Studio drivers via the NVIDIA App.
For teams comparing cloud render farms versus desk boxes, this continues the local AI factory thread — capital expense hardware instead of per-minute GPU rental, with Hetzner auction VPS as the cloud counterexample.
Microsoft Ignite cross-links in the same post
NVIDIA's November 18 article also summarizes Ignite 2026 items relevant to Spark buyers:
- Windows Server 2028 support coming to DGX Station for Windows on GB300 Grace Blackwell Ultra hardware.
- Windows ML can load GGUF models from Hugging Face via llama.cpp backend.
- DxCGC experimental DirectX 12 ML backend for game developers.
- Windows APIs gaining GPU support for AION Instruct and VSR.
Read our Ignite hub for license and GA nuance; this post only notes what NVIDIA repeated for Spark audiences.
OEM models named in the blog
NVIDIA calls out ASUS ProArt GR1X and MSI EdgeMesa N AI+ Series as examples. Dell is listed among available vendors without a specific SKU name in the excerpt we fetched. Acer, GIGABYTE, and HP remain coming soon, as does Microsoft Surface RTX Spark Dev Box.
Pricing is not in NVIDIA's post — check each OEM configurator. Compare against DGX Spark developer kit if you need lab-grade cooling and CUDA-Q adjacency instead of living-room form factors.
What to verify before you buy
- Confirm 128GB and GPU tier on the exact SKU; marketing names vary by region.
- Driver channel — Studio versus Game Ready — if you live in Blender and Premiere daily.
- Model catalog you plan to run: link Nemotron 3.5 checkpoints and DeepSeek NVFP4 builds you already test in cloud.
- Agent policies — local agents still need MCP and shell guardrails when agents touch git or package publish.
Common questions about RTX Spark desktops
Is RTX Spark the same as DGX Spark? NVIDIA positions RTX Spark compact PCs for creators and local agents, while DGX Spark is the developer-kit supercomputing line we covered at SC26. They share branding but target different buyers and software bundles.
Can I run OpenAI models locally on Spark? NVIDIA’s post emphasizes Hermes, Unsloth, ComfyUI, and Windows ML GGUF workflows — not ChatGPT weights. You run open or NVIDIA-partnered checkpoints from Hugging Face and similar hubs.
Do I need Windows on Arm? Spark desktops ship as Windows systems with Arm-native creative apps; verify each Adobe build’s beta status before you migrate a production edit suite.
Software versions named in NVIDIA’s post
NVIDIA’s November 18 article bundles several versioned dependencies worth pinning in your runbook:
- Blender 5.3 with DLSS 4.5 Ray Reconstruction in Cycles viewport.
- llama.cpp v1.0.0 with NVFP4 paths for local inference.
- Ripple-65B Preview and Runway Praxis-1 checkpoints on Hugging Face in NVFP4.
- TensorRT Edge-LLM walkthrough for Nemotron-3.5 Lightning on DGX Spark family hardware.
None of those replace your regression tests — diffusion and video world models drift from marketing FPS claims quickly. Capture before/after clips when you evaluate ComfyUI RTX Video nodes on Spark versus your existing 3090 or laptop baseline.
OEM SKUs and buyer personas
| Persona | Why Spark desktop | Caveat |
|---|---|---|
| Indie creator | ComfyUI + Premiere on one box | Verify Adobe Arm beta apps |
| Agent builder | Hermes always-on | Secure shell tools |
| Fine-tuner | 128GB for large LoRA | Still not H100 cluster |
| Enterprise dev | 10GbE to NAS | IT MDM may lag consumer launch |
Microsoft Surface RTX Spark Dev Box targets Windows ML developers testing GGUF flows announced at Ignite — overlap with Llama Windows tray app.
Competitive landscape: Mac Studio, DGX Spark, cloud
Mac Studio buyers compare unified memory versus CUDA ecosystem. Spark wins on NVIDIA stack depth (TensorRT, DLSS, CUDA-Q adjacency in other posts). Apple wins on per-watt for some creative tasks — this is not a benchmark post; your codecs decide.
DGX Spark remains the lab kit with SC26 messaging. Buy RTX Spark desktop when desk form factor and OEM warranty matter more than rack airflow.
Cloud break-even: if you spend less than ~$300/month on APIs, desktop amortization may lose to burst cloud unless privacy or offline is mandatory — rerun math with your kWh rate.
Power, thermals, and desk placement
Compact Grace + Blackwell boxes still draw datacenter-class power when 128GB unified memory feeds 120B-class loads. Plan circuit capacity, UPS, and acoustic tolerance — 10GbE helps only if NAS can saturate the link during checkpoint sync.
Always-on agents mean sleep states stay disabled. Compare electricity cost to cloud API spend using your tariff; AIRLLM math on small GPUs is not transferable, but break-even hours still matter for finance.
Related reading
- RTX Spark launch software ecosystem
- Llama Windows tray app
- PAIR and Hermes desktop
- DGX Spark developer kit SC26
- Microsoft Ignite 2026 hub
- AIRLLM and small-GPU inference
- Local model directories guide
Maintenance and driver hygiene
NVIDIA stresses Game Ready and Studio Driver channels through the NVIDIA App for Blender DLSS and Adobe TensorRT paths. Driver drift breaks reproducible ComfyUI graphs the same way CUDA version pins break training containers.
Create a quarterly driver + model manifest if you ship Spark boxes to distributed creators — PAIR one-click models reduce setup friction but not regression testing.
Ricursive Ripple and Runway Praxis checkpoints
NVIDIA highlights Ripple-65B Preview and Runway Praxis-1 as open or preview weights tuned for local Blackwell-class desktops. Both ship NVFP4 variants on Hugging Face per the November 18 post. Treat them as research toys until license and safety cards match your product tier.
World action models like Praxis-1 blur creative and robotics demos — useful for ScaleVLA readers comparing VLA stacks, but not a drop-in factory controller without sim validation.
When to wait for coming-soon OEMs
If your procurement cycle requires HP, Acer, or GIGABYTE service contracts, November 18 may be too early to standardize. NVIDIA’s post explicitly leaves those mini PCs as coming soon. Laptops remain available meanwhile — same Spark branding, different thermals.
Hardware claims and creative performance numbers come from NVIDIA's November 18, 2026 blog post.
