We independently research everything we recommend. When you buy through our links, we may earn a commission. Learn more ›
Hardware · Buying Guides
The Best Laptop for Running AI Locally
A 14-inch MacBook Pro with the M5 Pro chip and 64GB of unified memory is the best laptop for running AI locally for most people: it holds mid-size models comfortably and is supported by Ollama and LM Studio. If you need CUDA, get an RTX 5090 laptop such as the Razer Blade 16.
The Research Score rates each product on its own merits. Our picks also weigh price and who the product is for, so a pick can rank above a higher-scoring product. How we score
Running a large language model on your own laptop means your prompts and documents never leave the machine, you pay no per-token fees, and you can keep working offline. The catch is memory. A model has to fit into the memory your processor’s graphics can use, plus room for the conversation itself. That one number decides which models you can run far more than CPU speed or an “AI PC” sticker does.
For most people, the best balance is a 14-inch MacBook Pro with Apple’s M5 Pro chip and 64GB of unified memory. Its graphics can draw on most of that memory, so mid-size open models fit comfortably, and Ollama and LM Studio both support Apple silicon. If your work depends on NVIDIA’s CUDA software, a laptop with an RTX 5090 Laptop GPU and its 24GB of video memory is the fastest Windows route, but 24GB is a hard ceiling. If you want to run the biggest open models on a laptop, you’ll need 128GB of memory, from either an M5 Max MacBook Pro or an AMD Ryzen AI Max+ machine like HP’s ZBook Ultra G1a.
For this guide, we researched 13 laptops and laptop platforms for running large language models locally. We read manufacturer spec sheets and US store listings from Apple, Razer, HP, NVIDIA and AMD, the hardware documentation for Ollama and LM Studio, model sizes from Ollama’s library and OpenAI’s model card, and Notebookcheck’s review of the HP ZBook Ultra G1a, all checked in October 2026. We did not test any of these laptops or benchmark any models ourselves. Every price and spec below comes from a page we opened while researching, and those pages are listed in the sources. We rate each pick with our published Research Score. For laptops, features means how much memory is available to AI models and how fast it is, privacy means whether models run on the device, and compatibility means support for tools like Ollama, LM Studio and CUDA. How we research.
Who this is for
This guide is for people who want to run open models such as Llama, Qwen or OpenAI’s gpt-oss on their own laptop: developers, professionals who can’t send client documents to a cloud chatbot, and businesses that want AI without data leaving the device. If you only use ChatGPT, Claude or Gemini in a browser, any modern laptop will do.
How much memory you need
The most important question is how big a model you want to run. Models are usually shrunk (“quantized”) to about 4 bits per parameter for local use. AMD’s guide calls 4-bit the most common choice, and it’s what Ollama downloads by default. The download size is a good starting estimate for the memory a model needs. On top of that you need room for the conversation (the context), your operating system and your other apps.
Here’s what popular models take up, according to Ollama’s library and OpenAI’s model card:
Model
Download size (default)
What we’d plan for
Qwen3 8B
5.2GB
16GB laptop
Qwen3 14B
9.3GB
16GB to 24GB
gpt-oss-20b
runs within 16GB, per OpenAI
24GB
Qwen3 32B
20GB
32GB to 64GB, or a 24GB GPU with short context
Llama 3.3 70B (4-bit)
43GB
64GB minimum, 128GB to be comfortable
Llama 3.3 70B (8-bit)
75GB
128GB
gpt-oss-120b
fits a single 80GB GPU, per OpenAI
128GB
The right-hand column is our conservative planning estimate, not a measured requirement. Long documents and chat histories use more memory, so leave headroom.
Unified memory vs. GPU VRAM
Where that memory lives matters as much as how much you have.
Windows laptops with an NVIDIA GPU run models fastest when the whole model fits in the GPU’s own video memory (VRAM). NVIDIA lists 8GB on the RTX 5050 and 5060 Laptop GPUs, 8GB or 12GB on the RTX 5070, 12GB on the 5070 Ti, 16GB on the 5080 and 24GB on the 5090. A laptop’s system RAM doesn’t add to that.
Macs with Apple silicon use unified memory, a single pool shared by the CPU and GPU. A 64GB or 128GB MacBook Pro can put most of that toward a model.
AMD Ryzen AI Max+ laptops also share memory. AMD’s Variable Graphics Memory lets you assign up to 96GB of a 128GB system to the integrated GPU.
Memory bandwidth decides how fast the model answers once it’s loaded. Apple publishes it: 153GB/s for the M5, 307GB/s for the M5 Pro, and 460GB/s or 614GB/s for the two M5 Max versions. NVIDIA, Razer and HP don’t list it on the pages we checked.
Copilot+ PC NPUs vs. GPUs
Many Windows laptops are sold as “AI PCs” or Copilot+ PCs. Microsoft’s requirements for Copilot+ PCs are an NPU (neural processing unit) capable of 40 or more TOPS, 16GB of RAM and a 256GB SSD. But NPUs aren’t what runs most local chatbots: Ollama’s GPU documentation and LM Studio’s system requirements list GPUs and memory, not NPUs. A Copilot+ label with 16GB of RAM won’t get you far with local LLMs. Look at GPU memory first.
How we picked
Memory available to models: at least 24GB, with extra credit for 64GB and 128GB options.
Memory bandwidth: higher is faster.
Software support: Ollama, LM Studio, and CUDA where it matters.
Price for the configuration we recommend, from the manufacturer’s US store.
Portability, battery life and warranty, especially for business buyers.
We considered these laptops and platforms: Apple MacBook Air (M5), MacBook Pro 14 and 16 (M5, M5 Pro, M5 Max), Razer Blade 16 and Blade 18 (RTX 5090), ASUS ROG Strix SCAR 18 (RTX 5090), MSI Titan 18 HX (RTX 5090), HP ZBook Ultra G1a, ASUS ROG Flow Z13 (Ryzen AI Max+), and Copilot+ PCs with Snapdragon X, Intel Core Ultra 200V and AMD Ryzen AI 300 chips.
Top pick · The best laptop for running AI locally for most people
AM
Apple MacBook Pro 14 (M5 Pro, 64GB)
8.0/10
64GB of unified memory shared by the CPU and GPU, 307GB/s of memory bandwidth and native support in Ollama and LM Studio.
Apple’s unified memory is the big advantage here. A Windows laptop with a 16GB graphics card can keep only about 16GB of model on the GPU, while a MacBook Pro with 64GB of unified memory can put most of that toward the model, after leaving room for macOS and your apps.
Apple’s specs page lists the M5 Pro with up to 64GB of unified memory and 307GB/s of memory bandwidth. The configuration we recommend, with an 18-core CPU, 20-core GPU, 64GB of memory and a 1TB SSD, cost $3,699 at Apple’s US store at the time of research. That’s enough for the models most people actually run. In Ollama’s library, a 32-billion-parameter Qwen3 model is a 20GB download and OpenAI’s gpt-oss-20b is 14GB, so either fits with plenty of room left for long conversations and documents.
The software side is easy. Ollama’s documentation says it supports GPU acceleration on Apple devices through the Metal API, and LM Studio’s system requirements list Apple silicon Macs with 16GB or more of memory recommended. Apple rates the 14-inch model with M5 Pro at up to 22 hours of video streaming or 14 hours of wireless web use.
Flaws but not dealbreakers
Apple lists memory as a configure-to-order option, so choose the amount you’ll need for the life of the laptop when you buy. The M5 Pro also tops out at 64GB, which is tight for a 70-billion-parameter model. Ollama lists Llama 3.3 70B at 43GB in its default 4-bit form, which leaves limited headroom for context and other apps on a 64GB machine.
There’s no CUDA on a Mac. Most everyday tools for running models work fine, but if you fine-tune models or use research code written for NVIDIA GPUs, the Razer Blade 16 below is the safer choice.
Who it’s for
Developers, analysts and privacy-minded professionals who want to run mid-size open models on a portable laptop.
Who should skip it
Your tools require NVIDIA CUDA, or you need 70B-class models with long context, which call for 128GB.
Pros
64GB of memory the GPU can use for models
307GB/s memory bandwidth on this chip
Ollama uses the GPU through Apple's Metal
Up to 22 hours of video streaming, Apple says
Cons
$3,699 as configured
No CUDA; some research code assumes NVIDIA
Memory must be chosen when you order
Research Score
8.0/10
Features (30%)8
Value for money (25%)7
Privacy & security (20%)9
Compatibility (15%)8
Docs & support (10%)8
$3,699 at Apple as configured (18-core CPU, 20-core GPU, 64GB, 1TB SSD); 14-inch MacBook Pro with M5 Pro from $2,499 (at the time of research)
Affiliate link · price at the time of research, may have changed
Runner-up · Best Windows laptop for fast local AI with CUDA
RB
Razer Blade 16 (RTX 5090)
7.8/10
NVIDIA’s top laptop GPU with 24GB of GDDR7 video memory in a 16-inch chassis, for the widest software support.
If your workflow lives on NVIDIA’s software, buy a laptop with the GeForce RTX 5090 Laptop GPU. NVIDIA’s spec page lists it with 24GB of GDDR7 memory, more than any other RTX 50 laptop GPU. The RTX 5080 Laptop GPU has 16GB, and the RTX 5070 Ti has 12GB. NVIDIA’s page also points to LM Studio, Ollama and PyTorch as supported ways to run models, and Ollama’s GPU documentation lists the RTX 50 series among its supported cards.
We picked the Razer Blade 16 because it packs that GPU into what Razer calls its thinnest gaming laptop, and Razer says it runs the RTX 5090 at up to 165W. The configuration with an Intel Core Ultra 9 386H, 32GB of LPDDR5X system memory and a 2TB SSD was $4,899.99 on Razer’s US store at the time of research, and a 64GB option is available.
Ollama’s library lists Qwen3 32B at 20GB and gpt-oss-20b at 14GB, so both fit in 24GB, though the 32B model leaves little room for long context.
Flaws but not dealbreakers
Twenty-four gigabytes is the ceiling, and nothing you add to the laptop raises it. Llama 3.3 70B is a 43GB download even in Ollama’s default 4-bit form, so it won’t fit on the GPU. Software can spill part of a model into system memory, but that generally slows it down a lot. If big models matter more to you than CUDA, the 128GB picks below are a better fit.
It’s also an expensive way to buy model memory, and Razer notes that system memory is soldered and can’t be upgraded.
Who it’s for
Developers and tinkerers who need CUDA for tools like PyTorch, and who mostly run models of about 30B parameters or smaller.
Who should skip it
You want to run 70B-class models, or you want a quiet, long-lasting business laptop.
Pros
24GB of GDDR7 video memory, the most on any RTX 50 laptop GPU
CUDA works with nearly every local AI tool
RTX 5090 runs at up to 165W, Razer says
Cons
$4,899.99 as configured
24GB VRAM caps model size
System memory is soldered and not upgradeable
Research Score
7.8/10
Features (30%)8
Value for money (25%)6
Privacy & security (20%)9
Compatibility (15%)10
Docs & support (10%)6
$4,899.99 at Razer as configured (Intel Core Ultra 9 386H, RTX 5090, 32GB RAM, 2TB SSD; at the time of research)
Affiliate link · price at the time of research, may have changed
Budget pick · The cheapest laptop we'd buy for local AI
AM
Apple MacBook Air 13 (M5, 24GB)
7.6/10
A fanless, lightweight Mac whose 24GB of unified memory runs 8B to 14B models and gpt-oss-20b.
The MacBook Air is the least expensive laptop we’d recommend for local models, as long as you skip the base 16GB version. Apple’s US store listed the 13-inch model with the M5 chip, a 10-core GPU, 24GB of unified memory and a 512GB SSD at $1,499 at the time of research, and you can configure it with up to 32GB.
Twenty-four gigabytes is a practical floor for local AI. LM Studio recommends 16GB or more on Macs and says 8GB Macs should stick to small models. OpenAI’s model card says gpt-oss-20b can run within 16GB of memory, and smaller models such as 8B and 14B ones leave even more room. The Air runs the same Ollama and LM Studio software as the MacBook Pro.
Flaws but not dealbreakers
The M5 chip’s memory bandwidth is 153GB/s, according to Apple’s MacBook Pro specs, half that of the M5 Pro. In general, lower bandwidth means slower text generation with the same model. Apple doesn’t sell the Air with more than 32GB, so 30B-class models are a stretch.
Who it’s for
Students, writers and anyone who wants to try local models on an everyday laptop without spending pro-laptop money.
Who should skip it
You want to run 30B-class models, or you'll run long, heavy jobs that a fanless laptop isn't designed for.
Pros
$1,499 with 24GB and 512GB
Configurable to 32GB
Same Ollama and LM Studio support as the Pro
Cons
153GB/s bandwidth, half the M5 Pro's
24GB limits you to small and mid-size models
Research Score
7.6/10
Features (30%)6
Value for money (25%)8
Privacy & security (20%)9
Compatibility (15%)8
Docs & support (10%)8
$1,499 at Apple as configured (M5, 10-core GPU, 24GB, 512GB SSD); MacBook Air 13-inch from $1,299 (at the time of research)
Affiliate link · price at the time of research, may have changed
Upgrade pick · For the biggest models a laptop can run
AM
Apple MacBook Pro 14 (M5 Max, 128GB)
8.1/10
128GB of unified memory and 614GB/s of bandwidth, enough for 70B-class models and gpt-oss-120b.
If you want the largest open models on a laptop, memory is the only spec that matters, and the M5 Max has the most. Apple’s specs page lists the M5 Max with the 40-core GPU at up to 128GB of unified memory and 614GB/s of memory bandwidth, twice the M5 Pro’s. The 14-inch configuration with 128GB and a 2TB SSD cost $6,699 at Apple’s US store at the time of research.
That memory opens up a different class of model. Ollama’s library lists Llama 3.3 70B at 43GB in 4-bit and 75GB in 8-bit, and OpenAI’s gpt-oss-120b is designed to fit on a single 80GB GPU, according to its model card. Both fit in 128GB with room for context.
Flaws but not dealbreakers
The price is hard to justify unless you need models that don’t fit in 64GB. Apple’s specs say 128GB requires the 40-core GPU version of the M5 Max; the 32-core version tops out at 36GB. There’s no CUDA, and a full-precision 70B model, which Ollama lists at 141GB, still won’t fit.
Who it’s for
AI engineers and researchers who want 70B-class or 100B-plus mixture-of-experts models on a laptop.
Who should skip it
You mostly run models of 30B parameters or smaller, where the M5 Pro with 64GB costs $3,000 less.
Pros
128GB of unified memory
614GB/s memory bandwidth, the highest we found
Fits Llama 3.3 70B even at 8-bit
Cons
$6,699 as configured
128GB requires the 40-core GPU version of M5 Max
No CUDA
Research Score
8.1/10
Features (30%)10
Value for money (25%)5
Privacy & security (20%)9
Compatibility (15%)8
Docs & support (10%)8
$6,699 at Apple as configured (M5 Max with 18-core CPU and 40-core GPU, 128GB, 2TB SSD; at the time of research)
Affiliate link · price at the time of research, may have changed
Also great · Best Windows laptop for large models, and for business
HZ
HP ZBook Ultra G1a (Ryzen AI Max+ PRO 395, 128GB)
8.1/10
AMD’s Strix Halo chip lets its integrated GPU use up to 96GB of the laptop’s 128GB, in a 14-inch business workstation with a three-year warranty.
AMD’s Ryzen AI Max+ chips, code-named Strix Halo, bring Mac-style shared memory to Windows. AMD’s spec page lists the Ryzen AI Max+ PRO 395 with 16 CPU cores, Radeon 8060S graphics with 40 GPU cores, a 50-TOPS NPU, and support for up to 128GB of 256-bit LPDDR5X memory. AMD’s Variable Graphics Memory lets you assign up to 96GB of that to the GPU, and AMD says a 128GB system set that way can run 4-bit models of up to 128 billion parameters, or 16-bit models of up to 32 billion parameters.
HP’s ZBook Ultra G1a puts that chip in a 14-inch business workstation. HP’s US store listed the version with the Ryzen AI Max+ PRO 395, 128GB of memory, a 1TB SSD and a 14-inch WUXGA display at $4,999 at the time of research. The listing includes HP Wolf Pro Security and a three-year limited warranty covering parts and labor, without on-site repair. It runs Windows 11 Pro and qualifies as a Copilot+ PC.
Notebookcheck rated the ZBook Ultra G1a 14 87% and called its performance per volume “easily the strongest advantage of the model.” It praised the optional 120Hz OLED display and found sustained performance excellent for a 14-inch laptop.
Flaws but not dealbreakers
Software is the weak spot. Ollama’s GPU documentation lists the Ryzen AI Max chips only for Linux under AMD’s ROCm. On Windows, Ollama’s Vulkan backend is the documented route, and AMD’s own guides use LM Studio. Expect more setup and fewer compatible tools than with CUDA or a Mac. The memory bandwidth is also lower than a high-end Mac’s; HP and AMD don’t list a figure on the pages we checked.
Notebookcheck criticized the soldered RAM, average battery life, reduced performance on battery power and a relatively large 140W power adapter. HP’s store also showed an MSRP more than twice the sale price, so expect this laptop’s price to move around.
Who it’s for
Businesses and IT departments that need Windows 11 Pro, enterprise security and large local models on one machine.
Who should skip it
You need CUDA, or you want the fastest generation speeds on small models.
Pros
Up to 96GB of memory assignable to the GPU, AMD says
Three-year warranty in HP's listing
Windows 11 Pro and HP Wolf Pro Security
Cons
No CUDA; Ollama's AMD support for this chip is listed for Linux
Memory is soldered
Average battery life, per Notebookcheck
Research Score
8.1/10
Features (30%)9
Value for money (25%)7
Privacy & security (20%)9
Compatibility (15%)6
Docs & support (10%)9
$4,999 at HP as listed (Ryzen AI Max+ PRO 395, 128GB, 1TB SSD, 14-inch WUXGA display), shown as a sale price against an $11,856 MSRP (at the time of research)
Affiliate link · price at the time of research, may have changed
Other good options
If you want a bigger screen for an RTX 5090: Any laptop with the RTX 5090 Laptop GPU gets the same 24GB of video memory, according to NVIDIA, so an 18-inch model won’t raise the limit. We didn’t verify current US configurations and prices for 18-inch models such as the ASUS ROG Strix SCAR 18, Razer Blade 18 or MSI Titan 18 HX, so we didn’t make them picks.
If you want a 16-inch MacBook Pro: Apple sells the 16-inch with the same M5 Pro and M5 Max chips and memory options, starting at $2,999 at the time of research.
The competition
MacBook Pro 14 with the base M5: Up to 32GB and 153GB/s, like the MacBook Air, for more money.
RTX 5080 and 5070 Ti laptops: NVIDIA lists 16GB and 12GB of video memory. They’re fine for 8B to 14B models, but a 24GB Mac or a 5090 laptop has more room.
ASUS ROG Flow Z13: This tablet-style laptop is another Ryzen AI Max+ option. We focused on the ZBook Ultra G1a because we could verify a 128GB US configuration, its price and an independent review.
Copilot+ PCs with 16GB of RAM: They meet Microsoft’s requirements, but 16GB limits you to small models, as explained above.
What to look forward to
AMD says the iGPU in a 128GB Ryzen AI Max+ system can technically address up to 112GB of graphics memory, though it recommends staying within 96GB for best performance. We’ll update this guide as chips with more memory or bandwidth reach US stores.
FAQ
How much RAM do I need to run an LLM locally?
For small models (8B to 14B parameters), 16GB to 24GB is enough; LM Studio recommends at least 16GB. For 30B-class models, plan on 32GB to 64GB. For 70B-class models, you’ll want 64GB at minimum and 128GB for comfortable headroom. On a Windows laptop with a separate GPU, what counts is the GPU’s video memory, not system RAM.
Is a Copilot+ PC good for running local LLMs?
Not because of the Copilot+ label. Microsoft requires a 40-TOPS NPU and 16GB of RAM for Copilot+ PCs, but Ollama’s GPU documentation and LM Studio’s requirements don’t list NPUs as accelerators. For local chatbots, GPU memory matters far more.
Mac or NVIDIA laptop for local AI?
A Mac gives you much more memory for models at a given price (up to 128GB), while an NVIDIA laptop tops out at 24GB of video memory but supports CUDA. Choose NVIDIA if your tools need CUDA, and a Mac if you want bigger models.
Can a laptop run a 70B model?
Yes, with enough memory. Ollama lists Llama 3.3 70B at 43GB in its default 4-bit form, so it needs a laptop with 64GB or more of unified memory, or an AMD Ryzen AI Max+ laptop with 128GB. It won’t fit in the 24GB on any RTX 50 laptop GPU.
Is running AI locally safe for business data?
Running a model locally keeps prompts and files on the device instead of sending them to a cloud provider. You still need normal device security, such as disk encryption and access controls, and you should check each model’s license before using it commercially.
