New to this? Read it in simple words
  • Testers tried the first laptops with Nvidia’s RTX Spark chip.
  • They ran large AI models on the laptops themselves, without the internet.
  • Laptops with RTX 5090 graphics were faster on smaller models.
  • Some test laptops crashed or switched off when they got very hot.
Words to know
Local AI
An AI model that runs on your own computer, not on a company’s servers.
Parameters
The numbers inside an AI model that it learned in training; more usually means a bigger, more capable model.
Tokens a second
How fast a model writes; a token is a word or part of a word.

RTX Spark is Nvidia’s first chip for Windows laptops that puts the main processor and the graphics on one chip, sharing up to 128GB of memory. Microsoft’s Surface Laptop Ultra and HP’s OmniBook Ultra 16 are among the first laptops with it.

What the testers measured

Tom’s Hardware replayed AI agent tasks on three laptops with 128GB of memory, using the Qwen 3.6 35B model. The Surface Laptop Ultra took 8 minutes, Apple’s M5 Max MacBook Pro 10 minutes, and a laptop with AMD’s Ryzen AI Halo 13 minutes.

StorageReview had the Surface finish a coding task with the Codex tool and the same Qwen model in 88.6 seconds. It wrote up to 89 tokens a second.

The Register ran five models, from 26 billion to 122 billion parameters, on HP’s OmniBook Ultra 16. The biggest model wrote at about half the speed of the 35B one, but solved a coding task about 40 seconds sooner. Windows Central fitted models of 120 billion parameters entirely in the chip’s graphics memory.

Sources3457

Check
What independent testers found on the first RTX Spark laptops.

From reviews by Tom’s Hardware, StorageReview, The Register, PCWorld, Windows Central and The Verge on October 15 and 16. Each test used its own setup.

Where the chip falls short

Laptops with Nvidia’s RTX 5090 graphics were faster on models small enough to fit in their memory. On the Llama 3.1 8B model, a Razer Blade 16 wrote 128.8 tokens a second, against 56.4 for the Surface, StorageReview found. PCWorld measured 61.9 tokens a second for an RTX 5090 laptop, against 39.9 for a 32GB Surface.

Heat was a problem in long, heavy tests. A StorageReview Surface switched itself off three times at 95 °C, and a second one failed the same way. The Register’s HP laptop locked up above 90 °C, and PCWorld saw early app crashes, which Microsoft thought came from a faulty unit.

Several reviewers found gaming weak for the price.

Sources45679

Nvidia’s claims, and what costs what

Nvidia says the chip runs models of up to 120 billion parameters, with a context of 1 million tokens, and reaches “a petaflop” of AI computing. The testers confirmed that models of this size run. None of them tested the 1-million-token context or the petaflop figure.

The Surface Laptop Ultra costs from $2,599.99 to $5,899.99. HP’s OmniBook Ultra 16 starts at $3,199.99, and The Register’s 128GB test laptop cost $6,229.

Tom’s Hardware measured 10 hours and 43 minutes of battery life for a 32GB Surface. In The Verge’s test, an AI agent running a Qwen model on the laptop made a news briefing in 10 minutes and used 10% of the battery.

Sources125689

Sources

Every fact in this story comes from the sources below. Open them to check our work.

  1. 1
    Primary source · October 2026Slim Laptops & Small Desktops | NVIDIA RTX Spark Nvidia
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
How we checked this story

Every number comes from a named tester’s own measurements, each with its own setup, so numbers from different tests can’t be compared directly. Tom’s Hardware and Windows Central have the same owner.

More on this story

Mood: 6.2 out of 10 · Slightly positive

Independent tests confirm the laptops run 120-billion-parameter AI models locally, but heat shutdowns, crashes and prices above $5,000 temper the result. How we rate the mood

Basics explainedWhy do some huge models run cheaply? What do “reasoning” models do differently?
NvidiaLocal AIIndependent Testing