If you want to run AI models on your own machine and you need it to come with you, to a desk, a café, a flight, then you’re shopping for a laptop, and exactly one spec decides whether you’ll be happy: how much memory the model can live in. Not the CPU. Not the screen. Not even raw GPU horsepower. The single number that determines which models you can run at usable speed is the amount of fast memory the model fits inside, VRAM on an NVIDIA laptop, or unified memory on an Apple Silicon MacBook. Everything else in this guide flows from that one fact.
The honest short version: laptops can run local LLMs well, but they make real compromises a desktop doesn’t, and the best laptop for you is whichever one buys you the most usable memory within your budget, then survives sustained load without throttling itself into a slideshow. Let’s break down the two realistic paths, what each actually runs, and when you should buy a laptop at all.
Why memory is the deciding spec
A language model is a big pile of numbers (weights). To generate text quickly, those weights have to sit in fast memory the GPU can reach. If the whole model fits, you get good speed. If it doesn’t fit, the leftover spills into slower system RAM and runs partly on the CPU, and your speed collapses, often by an order of magnitude. There’s no graceful middle ground; it’s a cliff. If this distinction is new to you, our RAM vs VRAM explainer is the fastest way to internalize why the memory pool, not the clock speed, is the thing you’re really buying.
This is why two laptops with similar-sounding GPUs can have wildly different ceilings. A mobile GPU with 8GB and one with 16GB aren’t “a bit different”, they live in different model-size leagues. So before you compare anything else, find the memory number and treat it as the headline spec.
The two realistic paths
For local LLMs, laptop buying comes down to two camps that make opposite bets, the same split we cover in depth in Apple Silicon vs NVIDIA for local AI, just applied to portable hardware.
- NVIDIA gaming/workstation laptops bet on speed and software maturity. A discrete mobile GPU has high memory bandwidth and CUDA support, so the models that do fit run fast and “just work” with nearly every local AI tool. The catch is that mobile VRAM caps lower than desktop, and a thin chassis throttles under sustained load.
- Apple Silicon MacBook Pro bets on capacity and efficiency. Unified memory lets the GPU address a large shared pool, so a well-specced Mac can hold models a consumer laptop GPU can’t touch, quietly, on battery, without sounding like a jet. The catch is lower memory bandwidth than a top NVIDIA card (so big models run, but not blazingly), and a price that climbs steeply with memory.
There isn’t a universally “best” answer. There’s the right answer for what you want to run versus how fast you need it.
Path 1: NVIDIA gaming/workstation laptops
Here, the spec that matters is the mobile GPU’s dedicated VRAM. Mobile chips reuse desktop names but carry less memory than their desktop namesakes, and the exact figure is SKU-sensitive, so always confirm the specific laptop’s listed VRAM rather than trusting the model number. As a rough current map (verify per SKU before buying):
| Mobile GPU tier (laptop) | Typical VRAM | Reality for local LLMs |
|---|---|---|
| Entry discrete (e.g. xx70-class mobile) | ~8GB | Runs small models; tight. |
| Mid (e.g. xx70 Ti-class mobile) | ~12GB | Comfortable 7B-14B sweet spot. |
| Upper (e.g. xx80-class mobile) | ~16GB | More headroom, bigger context. |
| Top (current flagship mobile, e.g. 5090-class) | up to ~24GB | The mobile ceiling today. |
A few things to internalize from that table. First, the top mobile GPU today reaches roughly 24GB of VRAM, a real generational jump, but still where a mid-to-upper desktop card sits, not where the desktop flagships live. Mobile is always a tier or two behind on memory. Second, most gaming laptops people actually own land in the 8-16GB band, which is genuinely productive: that’s the range where the current crop of strong small-to-mid models shine. Our guides to the best local LLMs for 12-16GB VRAM and the best models for 8GB VRAM map specific models to those exact ceilings.
What an NVIDIA laptop realistically runs:
- ~8GB: 7B-8B-class models at a 4-bit quantization, with modest context. Fine for a capable everyday assistant; expect to keep context windows reasonable.
- ~12GB: the comfortable zone, 7B-14B models at 4-bit with room to breathe, or a 14B with shorter context.
- ~16GB: 14B comfortably, some 20B-24B-class models squeezed at tighter quantization/context.
- ~24GB: the mobile flagship, opens up larger 24B-32B-class models, though context headroom gets tight at the top.
On speed, the models that fit run fast on these cards thanks to bandwidth and CUDA, but pin numbers down only after you see them on the specific machine. A reasonable mental band for small-to-mid models on a healthy mobile GPU is in the tens of tokens per second, fast enough to feel conversational. Treat that as a ballpark, not a promise: thermals (below) can knock real-world figures down hard.
Path 2: Apple Silicon MacBook Pro
The MacBook Pro plays a different game. Instead of a fixed slab of VRAM, the Pro/Max chips share one unified memory pool between CPU and GPU, and the GPU can use most of it for a model. That means the memory tier you choose at checkout, not a fixed GPU spec, sets your ceiling, and the high tiers go far beyond any laptop GPU’s VRAM.
Configurations are generation- and chip-sensitive (Apple iterates the M-series yearly, and the Pro vs Max chip changes both bandwidth and the maximum memory you can order), so treat these as the shape of the lineup rather than fixed gospel:
- Pro-class chips typically top out in the mid range of unified memory (think on the order of ~48GB at the high end of that tier on current chips).
- Max-class chips can be ordered with much larger pools, historically up to around 128GB on the top configuration, with correspondingly higher memory bandwidth.
That capacity is the whole point. A MacBook Pro with a large-memory Max chip can hold 30B-class and even 70B-class models (at 4-bit, a 70B wants roughly 40-48GB of memory) that no single laptop GPU can fit at any usable speed. It does this quietly and on battery. The trade is bandwidth: even a Max chip’s memory bandwidth sits well below a top discrete NVIDIA GPU’s, so those big models run but generate at a more measured pace, perfectly usable for chat and writing, less ideal if you crave maximum tokens-per-second on small models. The base-config MacBook Pro (smaller memory, non-Max chip) is a weaker local-LLM machine than its price suggests; the value here lives in the memory upgrade, which is expensive.
One more honest note: Apple’s memory upgrades are not cheap, and you can’t add memory later, it’s soldered. Buy the capacity you’ll want in two years, because there’s no upgrade path.
The thermal and power reality (this is the catch)
Here’s the part the spec sheets won’t tell you: a laptop is a thermally and power-constrained box, and LLM inference is exactly the kind of sustained, heavy load that exposes those limits.
- Throttling under sustained load. Generating text is not a quick burst, it can hammer the GPU for minutes at a time. A thin laptop chassis can’t dissipate that heat indefinitely, so the GPU clocks down to stay cool. A desktop with the “same” GPU holds its peak far longer. Real-world laptop tok/s can drift below the headline figure once the machine heats up.
- Power limits. Mobile GPUs run at a lower power envelope than their desktop namesakes, and many laptops throttle further on battery versus plugged in. Plan to run heavy local-AI sessions on the charger.
- Fan noise and heat. Sustained inference means sustained fans. NVIDIA gaming laptops get loud and hot under this load; Apple Silicon is dramatically quieter and cooler for the same task, which is a genuine quality-of-life edge if you’ll work near other people.
- Battery life. Running a model locally drains a battery quickly. Portability gets you to the café; it won’t keep you running models there for long without a power outlet.
None of this makes laptops unusable, millions run local models on them happily. It just means you should size down your expectations versus a desktop with the same-sounding parts, and never assume the marketing tok/s figure survives a long session.
So which laptop should you buy?
A simple decision path:
- You mostly want small-to-mid models, fast, and you also game or do CUDA work: an NVIDIA laptop in the 12-16GB VRAM band is the sweet spot. It’s the most flexible all-rounder and runs today’s strong small models quickly.
- You want to run the biggest, smartest open models on the go and value silence/battery: a MacBook Pro with a large-memory Max chip. Nothing else portable holds 70B-class models. Pay for the memory, not the base config.
- Budget is tight and you just need a competent local assistant: an 8GB NVIDIA laptop (or a Mac with modest unified memory) runs capable small models. Set expectations to the 8GB model guide and keep context reasonable.
The honest part: should it be a laptop at all?
If you don’t truly need to carry it, a laptop is usually the worst value-per-dollar way to run local LLMs, and you deserve to hear that before you spend.
- A desktop gives you more memory, more speed, and no throttling for less money. Desktop GPUs carry more VRAM than their mobile namesakes, hold peak clocks indefinitely, and can be upgraded later. If you’re starting from scratch on a budget, a budget local-AI PC build will out-run a similarly priced laptop and keep doing it under sustained load.
- A mini-PC is the quiet middle path. A small always-on box on a shelf, covered in our best mini-PC for local AI guide, sips power, runs a private model 24/7, and several unified-memory mini-PCs now hold surprisingly large models for the money. It’s not portable, but it is small, cheap to run, and always there.
- The “best of both” move: run the model on a desktop or mini-PC at home and reach it from a cheap, light laptop over your own network. You get a thin, long-battery-life portable and the memory and speed of real desktop hardware, without paying the laptop tax twice.
Buy the laptop if portability is a hard requirement and you’ll genuinely run models away from home. If “portable” really means “it lives on my desk 95% of the time,” a desktop or mini-PC will make you happier and run bigger models for less.
Bottom line
The best laptop for local LLM work is the one that buys you the most usable fast memory inside your budget and holds up under sustained load. On the NVIDIA side, aim for the 12-16GB band for the best blend of speed and flexibility, knowing the mobile ceiling sits around 24GB today and that all of it can throttle when hot. On the Apple side, a large-memory MacBook Pro Max is the only portable machine that runs the truly big open models, quietly, on battery, if you’re willing to pay for the memory. And if you can live without portability, be honest with yourself: a desktop or mini-PC is almost always the better machine for the money. Either way, find the memory number first, everything else is detail.