Hands-on benchmarks of the MSI Prestige RTX Spark laptop with 128GB memory, tested against the DGX Spark on CPU, GPU, and thermal performance.
What is the MSI RTX Spark laptop? #
The MSI Prestige N16 Flip AI Plus is a 16 inch convertible laptop built around Nvidia’s RTX Spark platform, the laptop sibling of the DGX Spark desktop box. It pairs an Arm based N1X chip with a 20 core CPU and 128GB of LPDDR5 memory, running Windows on Arm. MSI markets it as capable of running 120 billion parameter language models locally without thermal throttling, a claim that independent testing against the DGX Spark largely backs up, with the laptop actually outperforming the desktop unit on several real world inference benchmarks.
TL;DR #
- The MSI Prestige N16 Flip AI Plus packs a 20 core Arm based N1X chip and 128GB of unified LPDDR5 memory into a 16 inch flip laptop with a 4K OLED touchscreen.
- In memory bandwidth tests , the laptop beat the DGX Spark, copying 8GB blocks at 143GB/s versus 113GB/s, which translated directly into faster token generation.
- Running GPT-OSS 120B through llama.cpp, the laptop hit 65 tokens per second compared to the DGX Spark’s 50.6, and performance barely dipped when unplugged.
- Windows on Arm software support has matured enough that native builds exist for Visual Studio, VS Code, Python, Node, and most Adobe apps, with only a handful of tools still running emulated.
- Switching from llama.cpp to vLM inside WSL boosted prompt processing from about 1,800 tokens per second to over 4,600, though single user generation speed dropped since vLLM is built for serving multiple concurrent requests.
- A 20 minute sustained load test with the CPU and GPU both maxed out showed no meaningful thermal throttling , with token generation staying flat around 50 to 52 tokens per second.
- The 140 watt charger is notably smaller than the DGX Spark’s power brick, and battery performance for GPU inference workloads stayed close to plugged in numbers.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does the RTX Spark laptop compare to the DGX Spark? #
The DGX Spark is Nvidia’s small form factor desktop box aimed at developers who want a local box with enough memory to run large models without a data center GPU. The RTX Spark platform brings similar specs (128GB of unified memory, a Blackwell class GPU) into laptop form, and MSI’s implementation is one of the first to ship.
Benchmarks run across CPU, GPU, and memory subsystems found the two machines trading blows depending on the workload. Raw GPU compute (matrix multiplication throughput) came out nearly identical: 101 on the laptop versus 97 on the DGX Spark. But memory bandwidth, which matters more for token generation than for prompt processing, favored the laptop by a wide margin: 143GB/s versus 113GB/s on writes, and 309GB/s versus 234GB/s on reads. That bandwidth edge showed up directly in real inference numbers, with the laptop generating tokens on GPT-OSS 120B about 30 percent faster than the DGX Spark using the same version of llama.cpp.
CPU heavy multi-threaded work, tested with a Mandelbrot set calculation across all 20 cores, told a more mixed story. The DGX Spark finished the Python version of the test in 16 seconds against the laptop’s 22 seconds running the same workload through Windows Subsystem for Linux. But isolating single core JavaScript performance showed identical speeds (1.24 seconds) on both Windows and Linux, suggesting the gap comes from how Python itself performs under each environment rather than a difference in raw chip speed.
Does Windows on Arm actually work for AI development? #
Yes, with a few caveats. Windows on Arm has been a rough experience historically, but Microsoft’s multi-year investment in native Arm tooling has paid off for most developer workflows. On this machine, Visual Studio, VS Code, the Zed editor, Python, and Node all run as native Arm binaries rather than emulated x86 code. Adobe’s Creative Cloud app, Photoshop, and Lightroom are also native Arm builds, though Premiere Pro still runs emulated.
Out of more than 250 background services and worker processes running on the system, only four were still x64, which gives a sense of how far Arm native support has come across the Windows ecosystem. For anyone worried about compatibility, the Works on WOA community tracking project is a useful reference for checking whether a specific tool runs natively.
The practical upshot: developers doing JavaScript or general tooling work can expect full native performance. Those doing heavy Python workloads should run them inside WSL (Windows Subsystem for Linux), since Python performance diverged noticeably between the two environments in testing, even though the underlying hardware was equally fast.
How fast can it run large local language models? #
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The laptop handled a range of model sizes without issue. A dense 27 billion parameter model (Qwen3) ran at 13.4 tokens per second, slower than the mixture of experts models because dense architectures activate every parameter on every token. GPT-OSS 120B, a mixture of experts model widely used as a benchmark despite its age, loaded in 28 seconds, used 61GB of the available 96GB allocated to the GPU, and generated at 65 tokens per second, a number that held steady even on battery power.
Pushing further, a newer 125 billion parameter model (Qwen3.8 Flash Next) loaded in 36 seconds, consumed 81GB of GPU memory, and still produced 33.7 tokens per second, a result MSI’s own marketing undersold given the company advertised 120 billion parameters as the ceiling.
Testing also compared llama.cpp (the engine behind tools like Ollama and LM Studio) against vLLM, which is what most production AI serving stacks use. vLLM isn’t natively supported on Windows but builds fine inside WSL. Prompt processing throughput jumped from around 1,800 tokens per second on llama.cpp to over 4,600 on vLLM, a 2.5x improvement. But for a single user generating text, vLLM actually fell behind, producing 35 tokens per second against llama.cpp’s 65, because vLLM is architected to serve many simultaneous requests efficiently rather than optimize for one user. With eight concurrent users, vLLM’s throughput rose to 127.5 tokens per second, which matters more for anyone running multiple AI agents at once than for a single chat session.
Does it throttle under sustained load? #
MSI’s claim of no thermal throttling held up in testing. Running the CPU across all 20 cores and the GPU simultaneously for 20 minutes showed token generation starting around 50 tokens per second and ending the test at 52, essentially flat performance with no meaningful degradation. The fan was audible under load but the system didn’t need to downclock to manage heat.
Unplugging the laptop had only a small effect on CPU heavy benchmarks (about a 5 percent slowdown on a multi-core compute test) and almost no effect on GPU inference speed, which stayed close to 65 tokens per second whether plugged in or running on battery. MSI’s companion software, MSI Center, offers Performance, Balanced, and Eco profiles, with Eco trading speed for a silent fan, giving users control over the tradeoff depending on whether they’re at a desk or on a plane.
Is the MSI RTX Spark laptop worth it for local AI work? #
For developers who want to run large local models without a desktop tower or a cloud bill, the combination of 128GB of memory, solid memory bandwidth, and no throttling under sustained load makes a reasonable case. It beat a DGX Spark on several real inference benchmarks while being portable enough to use unplugged, and Windows on Arm software support has matured to the point where most developer tools run natively. The main tradeoffs are Python performance being better inside WSL than native Windows, and vLLM’s Windows limitations requiring the same WSL workaround.
Frequently Asked Questions #
What CPU and memory does the MSI RTX Spark laptop have?
It uses an Arm based N1X chip with a 20 core CPU and 128GB of LPDDR5 unified memory, running Windows on Arm.
Is the MSI RTX Spark laptop faster than the DGX Spark?
On memory bandwidth and GPU token generation, yes, the laptop measured faster in testing. On some CPU heavy multi-threaded Python workloads, the DGX Spark came out ahead. Raw GPU compute throughput was roughly even between the two.
Remy doesn't write the code. It manages the agents who do. #
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Can it run 120 billion parameter models locally?
Yes. GPT-OSS 120B ran at 65 tokens per second using about 61GB of GPU memory, and a newer 125 billion parameter model ran at nearly 34 tokens per second while using 81GB.
Does performance drop when the laptop is unplugged?
CPU heavy multi-core workloads slowed slightly on battery, but GPU based AI inference speed stayed nearly identical whether plugged in or not.
Does Windows on Arm support developer tools well enough for AI work?
Most major developer tools, including Visual Studio, VS Code, Python, and Node, run as native Arm applications. A small number of tools, like Adobe Premiere Pro, still run emulated, and heavy Python workloads perform better inside WSL than natively on Windows.