Can NVIDIA RTX Spark Replace Your Dev Box for Local AI?

Running a serious model locally has meant one of two things: a desktop with a big GPU, or a Mac with a lot of memory. On October 16, NVIDIA added a third option. RTX Spark laptops went on sale from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599, and NVIDIA is pitching them straight at developers who want models and agents running on their own machine.

I haven't used one. This is the homework I'd do before spending that kind of money: what the chip is, what the memory number really buys you, and where the catches are. I just wrote up running a local coding model like Recursive Flint, and the first question there was the same one here. Does the model fit?

RTX Spark is a single superchip. It pairs an NVIDIA Blackwell RTX GPU with up to 6,144 CUDA cores and an NVIDIA Grace CPU with up to 20 cores, and the two share up to 128GB of unified memory. NVIDIA rates it at up to 1 petaflop of FP4 AI performance. MediaTek worked with NVIDIA on the CPU design.

The launch lineup includes the ASUS ProArt P14 and P16, Dell XPS 16, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and MSI's Prestige N16 and N16 Flip AI+. Compact desktops are due in November.

"Up to" is doing a lot of work in those specs. ASUS's ProArt laptops come in 24, 32, 48 or 64GB configurations as well as 128GB, and NVIDIA doesn't say which configuration the $2,599 price gets you. I'd assume the 128GB machines cost more.

Here's why 128GB matters. On a typical laptop or desktop GPU, the model has to fit in the card's own video memory, which is usually somewhere around 8 to 16GB on consumer cards. Go past that and you're offloading layers to system RAM, which is slow, or picking a smaller model.

Unified memory means the CPU and GPU draw from the same pool. On a 128GB machine, the GPU can use much more of that pool for model weights and context, minus whatever Windows and your other apps are holding. Flint's 23GB download is a good example: too big for a 16GB card, but it would leave plenty of room here.

ASUS puts rough numbers on it. Its guidance lists 7B to 14B models for 24 to 32GB configurations, 24B to 35B for 48 to 64GB, and 70B to 120B for 128GB, and it says the ProArt P16 supports models up to 120 billion parameters with a one-million-token context. ASUS also footnotes that this is illustrative, assumes quantization, and isn't validated performance. That's honest, and it's the right way to read it. Memory tells you what loads. It doesn't tell you how fast it runs, and nobody has published numbers I'd trust yet.

The Grace CPU is Arm, so these are Windows on Arm machines. For years that has meant "check whether your tools actually run." NVIDIA and Microsoft clearly know it. For launch, they list the first PyTorch-CUDA build for Windows on Arm, a native-by-default Python build, a native Conda installer, and support for 97% of the top 15,000 Python packages.

That last number is the one I'd think about. 97% is very good, and it also means some packages don't work yet. If your stack leans on a compiled dependency, an old x86-only binary or a niche CLI tool, check it before you buy.

The rest of the developer story is familiar. CUDA runs natively, so the same toolkit you'd use on a desktop GPU or in the cloud is there. Windows Subsystem for Linux is available for the Linux tools you'd use on a server. NVIDIA also lists day-one installers for agents including GitHub Copilot and Hermes, and a feature called NVIDIA PAIR that lets the machine serve inference to other Windows, Linux and macOS computers on your home network.

Buy one if you want to run models in the 30B to 120B range locally, you need code and context to stay on your own machine, and you want it all in a laptop you'd carry anyway. Agents that run tests and retry all day are a good fit too, since local runs don't bill by the token.

Keep your desktop GPU if your models fit in its memory now. A dedicated card in a tower has power and cooling no slim laptop can match, and you already own it.

Stay in the cloud if you need the largest frontier models, you train more than you run, or your local AI use comes in bursts. Paying per token for a few hours a week can beat $2,599 for a long time.

There are no independent benchmarks yet. The numbers out there are NVIDIA's and its partners'. Wait for real tokens-per-second tests on the models you use.

Some packages won't work. Windows on Arm support is much better now, but 97% isn't 100%.

The base price isn't the 128GB price. Price the configuration you'd actually need.

It's first-generation hardware. First-generation platforms tend to get their rough edges fixed in the second.

None of that rules it out. It just means you should spend a week on your own stack before you spend the money.

I'm not buying one this week. My current setup, in my AI coding setup for October 2026, is mostly cloud models, and the local experiments I'm running fit on hardware I already have. If I start running 70B-class models every day, or a client needs code that can't leave the building, RTX Spark goes on the short list.

If you're a small business trying to figure out what should run locally and what belongs in the cloud, that's work I take on. Get a free quote and tell me what you're running.

More memory on your desk is good news. Just make sure your tools can run on it too.

What is NVIDIA RTX Spark?

RTX Spark is NVIDIA's superchip for Windows PCs. It combines a Blackwell RTX GPU with up to 6,144 CUDA cores and a Grace CPU with up to 20 cores, sharing up to 128GB of unified memory, with up to 1 petaflop of FP4 AI performance.

How much do RTX Spark laptops cost?

NVIDIA says RTX Spark laptops start at $2,599, with pricing and configurations varying by region. Higher memory configurations, like 128GB, are likely to cost more.

Can RTX Spark run 70B or 120B models locally?

ASUS says its 128GB ProArt RTX Spark laptops can run selected 70B- and 120B-class models with appropriate quantization, and that context length and performance vary by model. There are no independent speed benchmarks yet.

Is RTX Spark Windows on Arm?

Yes. The Grace CPU is Arm-based, and NVIDIA's launch notes list the first PyTorch-CUDA build for Windows on Arm, native Python and Conda builds, and support for 97% of the top 15,000 Python packages. CUDA runs natively, and WSL is available for Linux tools.

Which laptops have RTX Spark?

At launch: the ASUS ProArt P14 and P16, Dell XPS 16, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and MSI Prestige N16 and N16 Flip AI+. Compact desktops are expected in November.

• NVIDIA Newsroom, "NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs"
• NVIDIA Blog, "NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs With RTX Spark and AI Agents"
• NVIDIA GeForce News, "RTX Spark Launches With Hundreds Of Top Games On October 16th"
• ASUS, "ProArt RTX Spark AI PC for Agents & Creativity"