Best NPU Laptops for Local AI: What to Buy

Buy the laptop by its memory ceiling, not by its NPU rating. Every Windows machine currently sold as a Copilot+ PC clears Microsoft's 40-TOPS bar, so the badge separates nothing — but the amount of memory the machine can be configured with separates everything, and on almost all of these laptops that memory is on-package or soldered, so the configuration you buy is the configuration you keep forever. Three machine classes can be ordered with 128 GB and will hold a 70B model: Apple's MacBook Pro with M5 Max (128 GB, up to 614 GB/s), Apple's M5 Pro with the 40-core GPU (128 GB, 307 GB/s), and AMD Ryzen AI Max+ "Strix Halo" systems (up to 128 GB unified). Every Intel Lunar Lake laptop, by contrast, is hard-capped at roughly 32 GB no matter what you spend, because the memory sits on the processor package.

How this page is ranked, before you read a single row. Everything below comes from published vendor and OEM specifications — Apple's tech specs, Qualcomm's platform pages, Intel ARK, AMD's Ryzen AI pages and Microsoft's own Copilot+ device list. This site does not own these laptops and has not benchmarked any of them. There are no tokens-per-second figures on this page, because we would have to invent them, and inventing them is exactly how hardware content goes bad. What you get instead is the set of things you can verify on a spec sheet before you spend the money: memory ceiling, memory bandwidth, whether a runtime exists that can reach that NPU at all, and where the price sits on a date we checked it.

If you want the silicon question — which chip has which NPU and how many TOPS it is rated at — that is a separate page: every CPU with an NPU, ranked by TOPS. This page is about picking a machine.

Which NPU Laptops Can You Actually Buy Right Now?

Sorted by maximum configurable memory, because that is the ranking that predicts what the machine can run. The "Where the ceiling comes from" column is the one most buyers miss.

Where the rows come from. The Snapdragon X machine list is Microsoft's own — the Copilot+ PCs developer guide names the Surface Laptop, Surface Pro, HP OmniBook X 14, Dell Latitude 7455, XPS 13 and Inspiron 14, Acer Swift 14 AI, Lenovo Yoga Slim 7x and ThinkPad T14s, Samsung Galaxy Book4 Edge, and ASUS Vivobook S 15 and ProArt PZ13, then adds that Copilot+ expanded to "AMD Ryzen AI 300 series and Intel Core Ultra 200V series". Apple's memory and bandwidth figures are from Apple's MacBook Pro tech specs and MacBook Air tech specs, read on 23 August 2026. Qualcomm's TOPS, memory and bandwidth figures are from the Snapdragon X2 Elite and Snapdragon X Plus platform pages. Intel's per-SKU NPU Peak TOPS (INT8) values are on Intel ARK.

Three things that table is quietly telling you:

• There is no 128 GB Intel laptop. Lunar Lake's memory is on the processor package, which is why every Core Ultra 200V machine — Dell, Lenovo, ASUS, Samsung, all of them — tops out in the same place. This is a silicon decision, not a product-segmentation decision, and no amount of budget gets around it.
• Qualcomm's 128 GB is a chip capability, not a shelf reality. Qualcomm publishes up to 128 GB of on-package LPDDR5X for the X2 Elite Extreme tier. Whether the specific laptop in front of you is offered in that configuration is a separate question you have to answer on the OEM's configurator, one SKU at a time.
• Apple is the only vendor where the highest-capacity and the highest-bandwidth option are the same order. 128 GB and 614 GB/s arrive together on the M5 Max with the 40-core GPU.

Why Is Memory the Spec That Decides This, Not TOPS?

Because generating a token means reading every active weight out of memory, once, per token. That makes decode a memory problem, not an arithmetic problem, and the NPU's TOPS rating describes arithmetic.

Here is the arithmetic, shown rather than asserted. A model quantised to Q4_K_M occupies roughly 0.6 GB per billion parameters. So:

• an 8B model is about 8 x 0.6 = 4.8 GB
• a 14B model is about 8.4 GB
• a 32B model is about 19.2 GB
• a 70B model is about 42 GB

Now subtract what the operating system and your open applications need — call it 8 GB on a Windows or macOS laptop, more if you keep a browser with forty tabs open. That gives a usable budget, and the budget is what decides which models you can load at all:

Bandwidth sets the second ceiling. Once the weights fit, tokens per second is capped by how fast the machine can read them. Divide the published bandwidth by the model size and you get an upper bound that no software can beat:

• A 4.8 GB (8B) model on a Snapdragon X Plus at Qualcomm's published 135 GB/s: 135 / 4.8 = ~28 tokens/second, ceiling.
• The same model on an M5 Max at Apple's published 614 GB/s: 614 / 4.8 = ~128 tokens/second, ceiling.
• A 42 GB (70B) model on that same M5 Max: 614 / 42 = ~14.6 tokens/second, ceiling.

Those are hard upper bounds from division, not measurements — real throughput is always lower because of attention, KV cache traffic and scheduling overhead. But they explain the shape of the whole category in one line: a bigger NPU rating never moves either ceiling. The mechanism is worked through properly in our guide to memory bandwidth and local LLM speed, and the model-to-memory mapping in more detail in the VRAM and unified-memory requirements table.

Which Laptop Has the Fastest NPU?

On vendor-published ratings, a Snapdragon X2 Elite laptop in one of the top three SKUs, at up to 85 TOPS. Nothing else currently sold is rated higher.

That is a genuine answer to the question people ask, and it is also close to useless as a purchase criterion, for two reasons worth being blunt about.

First, the numbers are not comparable. Intel labels its figure "NPU Peak TOPS (INT8)" on ARK. AMD quotes XDNA 2 at INT8. Qualcomm publishes 45, 80 and 85 on its own product pages with no precision stated anywhere on them. Apple has published no Neural Engine TOPS figure at all since M4. Ranking four vendors on a number that three of them define differently and one refuses to publish is not a ranking.

Second, nothing you are likely to run reaches the NPU. Ollama's own hardware documentation lists NVIDIA, AMD via ROCm, Apple Metal and Vulkan. There is no XDNA backend, no Intel NPU backend and no Hexagon backend. Getting a model onto an NPU means a vendor toolchain — OpenVINO on Intel, the QNN execution provider on Qualcomm, the Vitis AI execution provider on AMD — and a model converted for it. A rated 85 TOPS with no runtime pointed at it produces zero tokens per second.

So the honest ranking of "fastest NPU laptop" is: Snapdragon X2 Elite on paper, and it will not be the fastest laptop you can buy for running a local model. For that, look at the bandwidth column.

What Do These Laptops Cost?

Prices are the part of a hardware page that rots fastest, so this page prints only figures we actually looked up, with the date attached, and does not fill in the rest with plausible-looking numbers.

Checked on microsoft.com, 23 August 2026:

For every other machine in the table, check the OEM's own configurator on the day you buy. That is not a cop-out — laptop pricing moves weekly, previous-generation machines get discounted hard the moment a successor ships, and a stale price table on a buying guide is worse than no price table at all. What does not move is the memory ceiling, which is why this page ranks on that instead.

One pricing pattern that is stable enough to be worth stating: memory upgrades on soldered machines are the most expensive gigabytes you will ever buy, and they are also the only ones you can buy. On a desktop you defer the decision and add a DIMM later. On every laptop in the table above you cannot, so the upgrade you skip at checkout is an upgrade you have permanently declined. If your budget is fixed, take memory over the faster processor tier.

Can Any of These Laptops Actually Run a Model on the NPU?

Partly, and only through the vendor's own stack.

• Qualcomm ships the widest ready-made selection. Qualcomm AI Hub's Compute catalogue lists pre-converted models for Windows on Snapdragon — Qwen3 at 0.6B, 1.7B, 4B and 8B, Qwen3-VL, Gemma-4, Ministral-3 and GPT-OSS-20B among them — with runtime choices including ONNX Runtime, Qualcomm AI Runtime and a llama.cpp variant. These are converted, quantised builds, not ollama pull.
• Intel routes through OpenVINO, which exposes NPU as a device alongside CPU and GPU. It is a real path and it is a separate toolchain from anything you already use.
• AMD documents ONNX Runtime with the Vitis AI Execution Provider, which "intelligently determines what portions of the AI model should run on the NPU", plus OnnxRuntime GenAI and llama.cpp libraries for LLM applications (Ryzen AI Software documentation).
• Apple does not present the Neural Engine as an LLM target at all. On a Mac, local models run on the GPU through Metal or MLX, which is why Apple's bandwidth number matters and its Neural Engine number does not.

The practical read: if you want to type ollama run and have it work, you are choosing a machine on memory and bandwidth, and the NPU is along for the ride. If you specifically want to put a model on the NPU, Qualcomm currently has the shortest path from a standing start. A fuller version of the NPU-versus-GPU argument is in Copilot+ PC vs RTX for local AI.

Which One Should You Buy?

Matched to the constraint you actually have, using only things on the spec sheet.

You want to run a 70B model on a laptop. MacBook Pro with M5 Max and 128 GB, or M5 Pro with the 40-core GPU and 128 GB. Apple is the only vendor pairing that capacity with 307-614 GB/s of published bandwidth. Our Apple M5 guide for local AI goes deeper on which configuration is worth the money, and the Apple Silicon AI buying guide covers the same question across the range.

You want 70B-class capacity but need x86 and Windows. A Ryzen AI Max+ "Strix Halo" machine, configured with as much of the 128 GB as you can afford. Same silicon family as the mini-PCs in our Strix Halo mini-PC comparison, in a portable chassis. Verify the exact memory configuration on the OEM page before ordering — the ceiling is a chip capability and OEMs do not all offer the top tier.

You want a Windows-on-Arm machine and the highest rated NPU. A Snapdragon X2 Elite laptop. Check whether the specific SKU offers more than 16 GB before you commit, and be aware that x64 applications run under emulation.

You mainly want a good thin-and-light and will dabble in local AI. Any current Copilot+ machine will do, because they all clear the same bar. Spend the difference on the 32 GB option rather than the faster processor tier, and read the honest guide to running local AI on a laptop first so the expectations are calibrated before the box arrives.

You want maximum model size per dollar and portability is negotiable. Do not buy a laptop. A desktop with a used 24 GB card, or a mini-PC, will hold more model for less money and can be upgraded afterwards — the case is laid out in our budget local AI machine guide.

How Do I Check What a Specific Configuration Maxes Out At?

The configurator lies by omission — it shows you what that OEM chose to sell, not what the machine could take. Two checks before you order:

1. Find the exact processor part number, not the family. "Core Ultra 200V" spans 40, 47 and 48 TOPS. "Snapdragon X2 Elite" spans 80 and 85 with different memory tiers. Look the exact SKU up on Intel ARK or the matching Qualcomm or AMD product page.
2. Confirm whether the memory is on-package, soldered to the board, or in a socket. On-package (Lunar Lake, Snapdragon X, Apple) means the chip and memory ship as one part and the ceiling is absolute. Soldered-to-board is the same outcome for you as a buyer. A SODIMM or LPCAMM2 socket is the only case where a later upgrade exists, and it is rare in this category.

On a machine you already own, Get-CimInstance Win32_Processor | Select-Object Name in PowerShell gives you the exact part number to look up. On a Mac, the Apple menu → About This Mac is enough.

FAQ

Is a higher-TOPS laptop faster at running local AI?

Almost never, because local LLM decode is bound by memory bandwidth rather than arithmetic. Take two published figures: a Snapdragon X Plus at 135 GB/s and an Apple M5 Max at 614 GB/s. On a 4.8 GB model that is a ceiling of about 28 tokens/second against about 128 — a 4.5x gap — and the Snapdragon has the higher NPU rating of the two on paper. The TOPS number describes a block of silicon that most local AI software never touches.

Which laptop has the best NPU?

On vendor-published ratings, a Snapdragon X2 Elite laptop in one of the top SKUs, at up to 85 TOPS. AMD's Ryzen AI 400 series is the highest x86 rating at 60, and Intel's Panther Lake tops out around 50. Treat that as a spec-sheet ranking only: Qualcomm publishes no precision alongside its figure, Intel's is explicitly INT8, and Apple publishes nothing at all, so the four numbers are not measured the same way.

Do I need a Copilot+ PC to run local AI?

No. Copilot+ is a Windows feature gate, not a local-AI requirement. Ollama, LM Studio and llama.cpp run on machines with no NPU whatsoever and use the CPU or GPU. What a Copilot+ badge guarantees you is 16 GB of RAM and a 40+ TOPS NPU for Windows' own AI features — useful, but not the thing that decides whether a 32B model loads.

Can I add memory to an NPU laptop later?

On essentially none of the machines in the table above. Apple Silicon memory is part of the chip package. Intel Lunar Lake memory is on the processor package. Snapdragon X memory is on-package LPDDR5X. Strix Halo systems use soldered LPDDR5X. Choose the capacity at checkout, because there is no second chance, and prefer memory over a faster CPU tier if the budget forces the choice.

Is 16 GB enough for local AI?

For an 8B model at Q4_K_M, yes — 8 x 0.6 = 4.8 GB of weights leaves headroom on a 16 GB machine once the OS has taken its share. For anything larger it is not. A 14B needs about 8.4 GB and a 32B about 19.2 GB, and on a laptop that memory is shared with everything else you have open. 16 GB is the default Copilot+ configuration and it is the configuration most people later wish they had upgraded.

Why does an Intel AI laptop stop at 32 GB?

Because on Lunar Lake the memory is on the processor package rather than on the motherboard, so the capacity is fixed when the chip is manufactured. That is a silicon decision, and it applies identically across every OEM shipping that chip. If you need more than 32 GB in a Windows laptop, you are looking at Snapdragon X2 or Ryzen AI Max+, not at a more expensive Lunar Lake machine.

What This Page Does Not Claim

• We have not benchmarked any of these laptops. Every figure above is a vendor's or OEM's own published specification, reproduced with the label it came with, or arithmetic on those figures with the arithmetic shown. There are no tokens-per-second measurements here because we do not own the hardware, and a made-up number would be worse than no number.
• The tokens-per-second ceilings are division, not results. Bandwidth divided by model size is an upper bound. Real throughput is lower — sometimes much lower — depending on quantisation, context length, KV cache behaviour and which runtime you use.
• OEM configurations change constantly. A chip capability of 128 GB does not mean the laptop in front of you can be ordered that way. Verify on the configurator, on the day.
• Prices are only quoted where we checked them, with the retailer and date attached. Everywhere else the price column is deliberately empty.

Sources

• Microsoft — Copilot+ PCs developer guide (the 40-TOPS threshold and the named list of Copilot+ machines)
• Apple — MacBook Pro tech specs and MacBook Air tech specs (memory configurations and bandwidth, read 23 August 2026)
• Qualcomm platform pages: Snapdragon X2 Elite, Snapdragon X Plus
• Intel ARK — per-SKU NPU Peak TOPS (INT8) and memory specifications
• AMD Ryzen AI Software documentation — which runtimes reach the XDNA NPU
• Qualcomm AI Hub — Compute models — the pre-converted model catalogue for Windows on Snapdragon
• Microsoft Store listings for Surface Laptop, prices read 23 August 2026