Buying a laptop for local AI is different from buying one for gaming or office work. The brand of NPU sticker barely matters. What matters is how much memory the GPU can use and how fast that memory is, plus whether the laptop can hold full power for more than a minute. In this guide I explain those numbers in plain English, then give my picks for November 2026, from a sensible 16 GB RTX laptop to a 128 GB MacBook Pro that runs 120B-class models on a plane.
- Best overall for LLMs: MacBook Pro M5 Max (40-core GPU) with 64-128 GB unified memory, 614 GB/s. Runs 70B at Q4 and 120B-class MoE models. From $4,699 (14″) / $4,999 (16″).
- Best Windows/CUDA laptop: RTX 5090 laptop (24 GB GDDR7, 896 GB/s) with 150-175 W GPU power and vapor-chamber cooling. Best for ComfyUI and CUDA tools. About $3,200-$5,000+.
- Best big memory for less: ASUS ROG Flow Z13 with Ryzen AI Max+ 395, 64 GB (about $2,300-$2,400). Great for MoE models, slow for dense 70B.
- Best value: RTX 5080 (16 GB) or 5070 Ti (12 GB) laptop with 32 GB RAM for 8B-14B models and image generation.
- Skip: “AI laptops” sold on NPU TOPS with 16 GB of shared RAM and no real GPU.
What Matters in an AI Laptop
- Usable GPU memory. On Windows laptops this is the GPU’s VRAM (8-24 GB). On Macs and AMD Strix Halo machines it’s a large share of unified memory (36-128 GB). The model file plus context has to fit here to run fast.
- Memory bandwidth. Token generation speed is roughly bandwidth divided by model size. 896 GB/s (RTX 5080/5090 laptop) is about 3.5x faster than 256 GB/s (Ryzen AI Max) for the same model.
- Sustained power and cooling. Prompt processing and image generation hammer the GPU. Thin laptops with low GPU power limits throttle within a minute.
- Software. NVIDIA CUDA runs everything (ComfyUI, training, every custom node). Macs run LLMs superbly with MLX and llama.cpp, but image and video tools are slower and less complete. AMD is in between, see AMD GPUs for AI.
The NPU, the part marketed with TOPS numbers, runs Windows and camera AI features but does very little for local LLMs. I explain why in what is an NPU.
My Picks at a Glance
| Pick | Laptop | GPU memory / bandwidth | Price (Nov 2026, US) | Best for |
|---|---|---|---|---|
| Best overall (LLMs) | MacBook Pro 14/16 M5 Max, 40-core GPU | 64-128 GB unified / 614 GB/s | From $4,699 / $4,999 | 70B dense, 120B-class MoE, offline agents |
| Best Mac value | MacBook Pro 14 M5 Pro | Up to 64 GB / 307 GB/s | From $2,499 | 14B-32B models |
| Best Windows / CUDA | RTX 5090 laptop (16″, 150-175 W) | 24 GB GDDR7 / 896 GB/s | about $3,200-$5,000+ | ComfyUI, 27B-32B LLMs, training small LoRAs |
| Best value Windows | RTX 5080 or 5070 Ti laptop | 16 GB / 896 GB/s or 12 GB / 672 GB/s | varies | 8B-14B LLMs, image generation |
| Big memory, lower price | ASUS ROG Flow Z13, Ryzen AI Max+ 395, 64 GB | Up to about 48 GB usable / ~256 GB/s | about $2,300-$2,400 | MoE models, long context, portability |
| Budget | Any laptop with 32 GB RAM | CPU only / ~90 GB/s | low | Learning with 3B-8B models |
MacBook Pro M5 Max and M5 Pro
If your main goal is running big language models on a laptop, the MacBook Pro M5 Max is the default answer in 2026. The 40-core GPU version has 614 GB/s of memory bandwidth and up to 128 GB of unified memory, so a 70B model at Q4 (about 42.5 GB) fits with room for context. Testers have reported 64-88 tokens/s for gpt-oss 120B Q8 and 55-66 tokens/s for Qwen3.5-122B-A10B on a 128 GB M5 Max, numbers no other laptop matches for models that size.
- Watch the binning: the 32-core M5 Max has 460 GB/s and caps at 36 GB. Only the 40-core version reaches the 64-128 GB tiers.
- M5 Pro: half the bandwidth (307 GB/s) and up to 64 GB. Good for 14B-32B models, too slow for 70B.
- Battery: 99.6 Wh on the 16″, 72.4 Wh on the 14″. LLM inference on battery is practical, unlike on RTX laptops.
- Downsides: memory is soldered, ComfyUI and video models run slower than on RTX cards, and CUDA-only tools don’t run.
Use MLX-format models in LM Studio for the best speed on a Mac. My MLX vs GGUF guide explains when each is faster.
RTX 5090, 5080 and 5070 Ti Laptops
Laptop GPUs share names with desktop cards but not specs. The RTX 5090 laptop uses the GB203 chip (the desktop 5080’s chip) with 10,496 CUDA cores and 24 GB of GDDR7 at 896 GB/s, not the desktop 5090’s 32 GB.
| Laptop GPU | VRAM | Bandwidth | 8B Q4 ceiling (theoretical) | Fits fully on GPU |
|---|---|---|---|---|
| RTX 5090 laptop | 24 GB | 896 GB/s | about 183 tok/s | Up to about 27B-32B at Q4 with short context |
| RTX 5080 laptop | 16 GB | 896 GB/s | about 183 tok/s | Up to about 14B-20B |
| RTX 5070 Ti laptop | 12 GB | 672 GB/s | about 137 tok/s | Up to about 14B at Q4, short context |
| RTX 4090 laptop (used) | 16 GB | 576 GB/s | about 118 tok/s | Up to about 14B-20B |
Ceilings are bandwidth divided by a 4.9 GB model file; real speeds are lower. The key buying rule for RTX laptops: check the GPU power (TGP). A 5090 laptop at 150-175 W with a vapor chamber performs very differently from a thin 125 W model that throttles after 30-60 seconds of heavy prompt processing. Plan to run AI work plugged in.
For image and video generation, an RTX laptop is the best portable option because the whole ComfyUI ecosystem is built on CUDA.
Ryzen AI Max+ 395 Laptops (ROG Flow Z13)
AMD’s Strix Halo chip pairs a 40-CU Radeon iGPU with up to 128 GB of unified LPDDR5X at roughly 256 GB/s. That bandwidth is well below a MacBook M5 Max or an RTX laptop, but the memory pool is huge for the price.
- ASUS ROG Flow Z13: 32 GB about $2,070, 64 GB about $2,300-$2,400 (the one to buy), 128 GB about $2,800 mostly via resellers.
- Good at: mixture-of-experts models where only a few billion parameters are active per token, and long-context work.
- Weak at: dense 70B models (around 5 tokens/s) and heavy image or video generation.
- Tip: raise the GPU memory allocation in AMD Software or BIOS so LM Studio can use most of the RAM.
Other Strix Halo laptops (for example HP’s ZBook Ultra G1a) follow the same pattern. If you don’t need portability, a mini PC with the same chip or a DGX Spark-class box is cheaper per GB; see Mac vs PC for local AI.
Snapdragon X2, Intel Panther Lake and RTX Spark
- Snapdragon X2 Elite: 80 TOPS NPU, excellent battery life, but LLMs mostly run on CPU or NPU with limited model support. Fine for 3B-8B models.
- Intel Core Ultra Series 3 (Panther Lake): stronger Arc integrated graphics and about 50 TOPS NPU; good thin-and-light laptops for small models through OpenVINO or llama.cpp’s SYCL/Vulkan backends.
- NVIDIA RTX Spark (N1X) laptops: Arm Windows with a Blackwell GPU and up to 128 GB unified memory. Promising for local AI on paper; check independent reviews and software support for your tools before buying.
What Each Tier Runs
| Laptop tier | LLMs (comfortable) | Image generation | Video generation |
|---|---|---|---|
| 32 GB RAM, no dGPU | 3B-8B, slow | Barely | No |
| RTX 5070 Ti (12 GB) | 8B-14B | Yes, FP8/GGUF | Short clips with GGUF |
| RTX 5080 (16 GB) | 8B-20B, MoE with offload | Yes | Possible with GGUF |
| RTX 5090 (24 GB) | Up to 27B-32B | Excellent | Yes, best portable option |
| Ryzen AI Max+ 395, 64-128 GB | MoE up to 120B-class, dense 70B slowly | Slow | Not recommended |
| M5 Pro, 48-64 GB | 14B-32B | Moderate | Slow |
| M5 Max 40-core, 64-128 GB | 70B dense, 120B-class MoE | Moderate | Slow |
For exact model sizes, see how much VRAM do I need for AI.
Buying Checklist
- Decide your main job: LLMs (favour memory size) or image/video (favour CUDA).
- Check usable GPU memory: VRAM on RTX laptops, unified memory tier on Macs and Strix Halo.
- Check bandwidth: 600+ GB/s for fast 30B+ models.
- On RTX laptops, check GPU TGP (150 W+) and cooling reviews.
- Get at least 32 GB of system RAM on Windows laptops (64 GB if you’ll offload MoE layers).
- Get 2 TB of storage; models fill drives fast.
- Memory is soldered on most 2026 AI laptops, and RAM prices are high, so buy the memory you’ll need for 3-4 years now. Background in why RAM is so expensive in 2026.
Laptop vs Desktop vs Cloud
A desktop with the same money gets you more: a desktop RTX 5090 has 32 GB against the laptop 5090’s 24 GB, desktop cards run at full power with better cooling, and you can add a second GPU later. If you only need big models occasionally, renting is cheaper still; see how to run ComfyUI on RunPod and cheapest cloud GPUs. Buy a laptop for local AI when you truly need it offline or on the move.
Getting the Most From an AI Laptop
The model runs slower than expected
On Windows, check the dedicated GPU is in use (Task Manager > Performance > GPU 1) and that the laptop is in its performance mode, plugged in. Set the NVIDIA Control Panel to use the high-performance GPU for LM Studio, Ollama and Python.
Ollama uses the integrated GPU or CPU
Run ollama ps. If it isn’t 100% GPU, update the NVIDIA driver and disable hybrid “Eco” modes. My VRAM checking guide covers the commands.
The Mac won’t load a model that should fit
macOS limits how much unified memory the GPU can wire. LM Studio shows the limit; advanced users can raise it temporarily with sudo sysctl iogpu.wired_limit_mb= followed by a value, leaving at least 8 GB for macOS.
Fans are loud and speed drops after a minute
That’s thermal throttling. Raise the laptop on a stand, use the manufacturer’s performance profile, or cap power slightly for steadier speeds.
Strix Halo laptop shows only 16 GB of GPU memory
Increase the GPU memory allocation (UMA frame buffer) in AMD Software or the BIOS.
Frequently Asked Questions
What is the best laptop for running local LLMs in 2026?
A MacBook Pro with the 40-core M5 Max and 64-128 GB of unified memory. It has 614 GB/s of bandwidth and can run 70B models at Q4 and 120B-class MoE models.
Is an RTX 5090 laptop good for AI?
Yes, it’s the best Windows laptop for AI, with 24 GB of GDDR7 at 896 GB/s and full CUDA support. Buy a model with 150-175 W GPU power and strong cooling, and use it plugged in.
How much RAM do I need in a laptop for local AI?
At least 32 GB on a Windows laptop with a dedicated GPU, 64 GB if you offload large MoE models. On a Mac or Strix Halo laptop, the unified memory is your GPU memory, so 48-64 GB is a good minimum for serious use.
Does the NPU in an AI laptop help run LLMs?
Only a little. NPUs run Windows and camera AI features well, but most local LLM tools use the GPU or CPU. Memory size and bandwidth matter far more.
Is the ROG Flow Z13 good for local LLMs?
The 64 GB Ryzen AI Max+ 395 version is good for MoE models and long context at a reasonable price. It’s slow for dense 70B models and for image or video generation.
Is a MacBook or an RTX laptop better for Stable Diffusion and ComfyUI?
An RTX laptop. ComfyUI and most image and video tools are built around CUDA, so they run faster and with fewer compatibility issues on NVIDIA.
Can a laptop run a 70B model?
Yes, a MacBook Pro M5 Max with 64 GB or more runs 70B at Q4 at usable speeds. Ryzen AI Max+ laptops with 64-128 GB can load it but run slowly. RTX laptops with 24 GB need heavy offloading.
Final Thoughts
Pick by workload. For big LLMs on the go, nothing beats an M5 Max MacBook Pro with lots of memory. For ComfyUI, image and video work, get an RTX laptop with a high power limit, ideally a 5090 or 5080. If you want a lot of memory without the Apple price, a 64 GB Ryzen AI Max+ 395 machine is the clever middle ground. And if you can work at a desk, a desktop or a rented GPU gives you far more for the money.
Related: best GPU for local LLMs, RTX 5090 vs 5080 for AI and the best Ollama models by VRAM.
Sources
- Local AI Master: Best laptop for local AI, what each tier runs (updated 2 Nov 2026)
- OpenClaw DC: MacBook Pro vs RTX 5090 vs Strix Halo for local LLMs (14 Aug 2026)
- Apple: MacBook Pro technical specifications
- NVIDIA: GeForce RTX 50 series laptops
- ComputingForGeeks: Best laptops for local LLMs, VRAM vs unified memory