RTX Spark for coding agents: what Nvidia shipped

RTX Spark is Nvidia's chip for Windows laptops and compact desktops: a Blackwell RTX GPU, a 20-core Grace CPU and up to 128 GB of unified memory, rated at one petaflop of FP4. Laptops shipped on October 16, 2026, and compact desktops go on sale in November. For coding agents, the memory is what matters. 128 GB holds models that no single consumer graphics card can.

#What Nvidia and Microsoft announced

Nvidia and Microsoft announced RTX Spark on October 7, 2026. The headline specs:

• GPU: Nvidia Blackwell RTX, up to 6,144 cores.
• CPU: Nvidia Grace, up to 20 cores, linked to the GPU at 600 GB/s.
• Memory: up to 128 GB, shared by the CPU and GPU.
• AI performance: one petaflop at FP4.

Laptop preorders opened the same day. Compact desktops follow in November. Nvidia lists Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte as partners. Microsoft's device is the Surface Laptop Ultra.

Nvidia has not published prices or memory bandwidth in that announcement. Both matter for a buying decision, so wait for reviews before you compare it on tokens per second.

#Why unified memory matters for agents

A coding agent's model has to fit in memory along with its context. Long sessions build up a lot of context: files read, tool output, diffs.

On a regular PC, the model has to fit in the graphics card's memory. The largest consumer card, the RTX 5090, has 32 GB. That caps you at 30B-class models with comfortable context.

With unified memory, the GPU can use most of the system's RAM. At 128 GB, a 70B model at 4-bit quantization fits with plenty of room for context. Some larger mixture-of-experts models fit too.

This is the same design as Apple silicon Macs, AMD's Strix Halo and Nvidia's own DGX Spark. RTX Spark brings it to Windows with the full CUDA stack. Our mini PC comparison covers the other three.

#The CPU is Arm

Grace is an Arm CPU, so RTX Spark machines run Windows on Arm. Most developer tools now ship Arm builds. Some older x86-only tools will run under emulation, more slowly, or not at all.

Check your toolchain before you buy: compilers, language runtimes, Docker images and any native dependencies your projects build. Docker images in particular need an arm64 variant.

#Microsoft Execution Containers

The same announcement made Microsoft Execution Containers (MXC) generally available. Nvidia describes MXC as "the OS-level infrastructure that lets agents run safely and persistently in the background, under operating system control."

That's the right direction. A coding agent with shell access needs a boundary around what it can touch. It needs to keep running when you close the terminal. Building this into the OS means Windows apps can rely on it.

The details will decide how useful it is for coding agents. Look for three things as Microsoft documents it:

1. Filesystem scope. Can you give the agent one repository and nothing else?
2. Network control. Can you deny outbound traffic by default and allow specific hosts?
3. Credential handling. Does the agent see your real API keys, or a token it can't take with it?

These are the questions any agent sandbox has to answer. We wrote up the full list in what a coding agent sandbox really needs.

#Faster local inference

Nvidia also announced inference speedups at IFA 2026 in September. The numbers are specific to the hardware they were measured on:

• llama.cpp: up to 1.9x higher throughput on a GeForce RTX 5090.
• vLLM: 1.2x on an RTX PRO 6000 Blackwell Workstation Edition, and up to 1.4x across two DGX Spark units.

Nvidia didn't state the baselines. Treat these as Nvidia's own claims, not results you'll see on an RTX Spark laptop.

The same post announced NVIDIA PAIR, a free open-source beta that finds PCs on your local network and routes inference between them. It works with Ollama and LM Studio. That's useful if you want one machine to serve models to agents running somewhere else.

#Running coding agents on RTX Spark

There are two practical setups.

Use it as the model server. Run Ollama or LM Studio on the RTX Spark machine and point your agent at it. Autopilot supports both as providers, alongside 25+ API providers using your own key. You can use the local model for cheap steps and switch to a frontier model when a task needs more.

Use it as your laptop, with agents on a server. dl installs on Windows, Mac or Linux. Agents that run on this computer need Mac or Linux, so on Windows dl runs them on your own server over SSH. They keep going while the laptop is closed. Each task runs in a Hotcell sandbox, and real API keys stay outside it.

See the install guide to set it up.

#Who should consider it

RTX Spark makes sense if you work on Windows, want 70B-class models locally, and your toolchain runs on Arm. It also suits anyone who needs CUDA specifically, which Apple silicon and Strix Halo don't offer.

If you're on Linux or macOS already, compare it with DGX Spark, a Mac with enough memory, or a Strix Halo box. Wait for independent numbers on memory bandwidth and sustained throughput first. For agents that run for hours, steady output matters more than peak.

#FAQs

#What is Nvidia RTX Spark?

A chip for Windows laptops and compact desktops that combines a Blackwell RTX GPU (up to 6,144 cores), a 20-core Grace CPU and up to 128 GB of unified memory. Nvidia rates it at one petaflop of FP4.

#When can I buy an RTX Spark PC?

Laptops have been available since October 16, 2026. Compact desktops go on sale in November 2026. Partners include Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte.

#Does RTX Spark run x86 Windows apps?

It runs Windows on Arm, because Grace is an Arm CPU. Many x86 apps run under emulation, but check your developer tools and Docker images for arm64 support first.

#Can RTX Spark run a 70B model?

With 128 GB of unified memory, a 70B model at 4-bit quantization fits with room for context. Speed depends on memory bandwidth, which Nvidia hasn't published.

#What is Microsoft Execution Containers?

MXC is Windows infrastructure that lets AI agents run in the background under operating system control. Microsoft made it generally available alongside RTX Spark in October 2026.

#How is RTX Spark different from DGX Spark?

Both pair a Grace CPU with a Blackwell GPU and up to 128 GB of unified memory. RTX Spark is for Windows laptops and desktops from PC makers. DGX Spark is Nvidia's own Linux mini PC.