Unplug it and let a big local model write your code — reading the "no outlet" reports from RTX Spark laptops

Starting at 70% battery in the morning and down to 31% by the evening: that is what the developer netrunner (@plotarmordev) wrote in a reply, adding "really not bad" ¹.

The work was getting NVIDIA's big Nemotron 3.5 Bolt model running on a Windows laptop. The post says it will "soon" run natively there, on battery ². A large model writing code on your own machine, with no hunt for an outlet: here is one developer's report from the doorway of that idea.

What was put together

According to the post, what runs is TensorFold on an RTX Spark N1X ^{3 4}. Add @MiaAI_lab's recipe and NVIDIA's LiLiCorr drafter, and the developer says code comes out up to 2.8x faster, at 100 tok/s ⁵. The developer says the port and a bug fix are up as PRs ⁶.

The parts, one by one:

• Nemotron 3.5 Bolt: an NVIDIA model with 74B total parameters, 9B of them active [[F404]]. It is described as designed to fit in 24GB with a native 2-bit quantization and run alongside other applications [[F403]]
• LiLiCorr: not a standalone language model but a drafter. It proposes draft token candidates, and the main model verifies them and produces the text [[F409]]
• TensorFold: an Apache-2.0 inference engine for running language models locally [[F400]] [[F401]]
• Mia's recipe: a recipe for serving the model on DGX Spark [[F411]]

So this is a DGX Spark recipe carried over to a Windows on Arm laptop. In the developer's PR, the recipe was built and served natively on Windows 11 ARM64, on an RTX Spark N1X laptop ⁷, without Docker or WSL ⁸.

How to read the benchmark table

The numbers come from a table in the recipe's Windows guide. The laptop ran on battery in Windows' Balanced power plan ⁹. Each figure is a single run, not a median ⁹, with thinking off and 512-token replies ¹⁰.

For one stream, TensorFold alone versus the recipe ¹⁰:

• Prose (greedy): 39.3 tok/s, then 59.9 tok/s [[F417]]
• Code (greedy): 36.4 tok/s, then 100.1 tok/s [[F417]]
• Prose (sampled): 36.6 tok/s, then 47.6 tok/s [[F417]]

With 4 streams at once, 91.7 tok/s became 117.6 tok/s (41.9 a stream) ¹¹. Loading took 66 seconds ¹², and with these settings the GPU held 33.2GB, the display's share included ¹³.

The post's "up to" multiplier is the developer's own phrasing for results under these conditions ⁵. The table is best read as the result of this procedure on this machine.

A few limits matter when you read it:

• It was checked on a single machine: an RTX Spark N1X laptop with 64GB, of which 40GB went to the GPU [[F413]]
• Only plain chat was checked. Images, video and structured output were not built [[F421]]

The post itself says the model will "soon" run natively on a Windows laptop ². The developer's reply collects the links to the recipe and the PRs ¹⁴.

The RTX Spark laptop as a thing you carry

If you imagine working that way, what kind of laptop lands on your desk? Gizmodo's Kyle Barr reviewed a higher-end configuration of the Lenovo Yoga 9n 2-in-1 hands-on ¹⁵.

According to Barr, it is the first laptop Barr had seen where a stylus works on both the screen and the trackpad ¹⁵, and the pen supports tilt and up to 8,192 levels of pressure sensitivity ¹⁶. Picture drawing on the screen with a pen while a local model runs in the background.

On power, Barr says the secret sauce of this chip is the promise of practically equivalent performance whether or not it is plugged in ¹⁷. That is a promise, not a measured result. On the review unit, Barr reports the following:

• Noise: even at top GPU load, the fan does not rise above a library whisper [[F425]]
• Trade-off: a max TDP of 70W, against 110W for the Surface Laptop Ultra [[F426]]. In 3DMark Speed Way it trailed by around 13%, and in Port Royal the gap was closer to 19% [[F427]]
• Memory: GPU memory defaults to 32GB and could be pushed to 48GB, which changed the benchmarks little [[F428]]
• Model size: with 64GB, mid-range models like Qwen 3.8 27B run at relatively fast speeds, enough for some agentic workflows in the background [[F429]]. For larger models you would still want 128GB [[F430]]
• Battery: a little more than 8 hours of browser work in best power efficiency mode at moderate brightness [[F443]]. Pushing the PC harder, closer to 6 or 7 hours, maybe less; with games or GPU load, 3 hours at best [[F442]]

The last item shows how much battery life depends on what you do. netrunner's "70% in the morning, 31% in the evening" is a separate report, from the developer's own reply, not from Barr's review unit ¹.

Other outlets measured other machines. Mobile01, testing the top MSI Prestige N16 Flip AI+ configuration, found no large gap between plugged-in and battery performance, but advises staying plugged in if you do a lot of AI inference or rendering ¹⁸. ITmedia PC USER measured video playback on an ASUS ProArt P14 at 100% brightness in the Balanced power plan, and the battery fell 7% in an hour ¹⁹. At that pace, ITmedia calculates about 13 hours 34 minutes, give or take an hour ²⁰. Both are results for the machine and conditions tested; the ITmedia figure is video playback, not an AI workload.

A third kind of Windows PC

A columnist at ASCII, writing in the Windows Info column (in Japanese), takes a wider view.

According to the post (in Japanese), Microsoft announced the Surface Laptop Ultra with RTX Spark, and the question is how a PC that can handle large language models will change Windows and PCs, smartphones and tablets ²¹. The columnist's reading:

• Until now, Arm machines running Windows were limited to Qualcomm's Snapdragon series [[F434]]
• Windows 11 PCs become 3 kinds: non-AI PCs, Copilot+ PCs and Spark-class PCs [[F435]]
• Copilot+ PCs handle small or mid-size language models; Spark-class PCs have the specs to handle large ones [[F436]]
• They will not be accepted widely right away, but for developers and creators they offer performance nothing else does, so some demand should exist [[F437]]
• If cloud AI costs swell, companies with PCs, smartphones and tablets will move to local inference [[F438]]

That is the columnist's opinion. Read after this, the post looks a little different: a large model on your own machine, away from the outlet.

If it goes on your desk (and in your bag)

NVIDIA's product page says that running locally means no API bills and data that stays on your machine ²². NVIDIA's blog also showed AI agents running on RTX Spark using Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B ^{23 24}.

What to check before buying, if you want big models away from an outlet:

• What you want to load: PC Watch, reviewing the 64GB ASUS ProArt P16, says 64GB is worth considering if you switch among quantized LLMs that fit, and 128GB if you expect more models or concurrent jobs [[F196]]. Barr also points to 128GB for larger models [[F430]]
• App support: Barr lists minor bugs such as stuttering in browsers [[F431]]. Call of Duty crashed because these laptops still await support from the Ricochet anti-cheat system, which NVIDIA lists as "coming later" [[F432]]. InDesign and Illustrator run through Microsoft's Prism emulator [[F433]]
• Specialist software and dev tools: Mobile01 strongly advises checking native support and operation before buying [[F231]]

Check the official support pages for the tools you want to run, on that machine and that OS, before you buy.

If you carry model files around, an external drive helps. Nemotron 3.5 Bolt is designed to fit in 24GB ²⁵, but if you swap several models while you experiment, a separate place to keep them is handy.

You can also look for RTX Spark laptops on Amazon.

Read next

On how much a local agent should be allowed to touch, see How much should a local AI agent be allowed to touch?. If memory size is the question, read a comparison of reviewers' memory measurements. For games and 3D, see How far do games and 3D go on RTX Spark laptops?.