Memory decides which local model a laptop can run, and speed comes second. No GeForce laptop ships with more than 24GB of VRAM, which stops you at a 32B model. A 128GB MacBook Pro or Strix Halo laptop loads 70B dense models and gpt-oss-120b, but decodes slower per gigabyte. So pick the memory pool first, then the chassis that keeps it fed.
We sized four laptops against real model files and published llama.cpp numbers, and every one of them is in stock on Amazon US. Prices and stock were re-checked on Amazon on November 6, 2026, three weeks before Black Friday, and that check moved the 128GB MacBook pick to the 14-inch model, which costs about $1,000 less than the cheapest 16-inch 128GB listing and ships from normal stock.
Pick a laptop by the model you run
Best for 70B models and 120B MoE: the MacBook Pro 14 with M5 Max and 128GB, $6,699 from Expercom, an Apple Premier Partner. It is the only laptop here that runs a dense 70B model at a usable speed, and it does it on battery.
Best value for big models: the ASUS ROG Flow Z13 with 128GB, $3,499 from HIDevolution. You get the same capacity class as the Mac for about half the money, and it is the best Linux machine in the group.
Fastest for models up to 32B: the ASUS ROG Strix SCAR 18 with an RTX 5090, $4,499.99 and sold by Amazon itself. It decodes anything that fits in 24GB faster than any Mac, and CUDA runs every training and serving tool you will want.
Best for 14B models and gpt-oss-20b: the MSI Vector 16 HX AI with an RTX 5080, $2,729 from a third-party seller. The 16GB card keeps the full 896 GB/s bus of the 5090, so small models run at nearly the same speed for $1,770 less.
Check the specs and prices side by side
The table carries the two numbers that decide an LLM laptop, the memory pool the model can use and the bandwidth that feeds it. The last column links each listing at the price we saw on November 6.
| Laptop | Memory for models | Bandwidth | Runs well | Buy on Amazon |
|---|---|---|---|---|
| MacBook Pro 14, M5 Max 40-core GPU, 128GB | about 107GB by default | 614 GB/s | 70B dense, gpt-oss-120b | $6,699 |
| ASUS ROG Flow Z13, Ryzen AI Max+ 395, 128GB | 96GB on Windows, about 108GB on Linux | 256 GB/s | gpt-oss-120b, 70B slowly | $3,499 |
| ASUS ROG Strix SCAR 18, RTX 5090 | 24GB GDDR7 | 896 GB/s | up to 32B | $4,499.99 |
| MSI Vector 16 HX AI, RTX 5080 | 16GB GDDR7 | 896 GB/s | up to 14B, gpt-oss-20b | $2,729 |
| MSI Vector 16 HX AI, RTX 5070 Ti (budget) | 12GB GDDR7 | 672 GB/s | 8B to 14B | $1,599 |
Only the SCAR 18 is sold by Amazon itself. The others ship from third-party sellers that Amazon lists in the buy box, and the Flow Z13 128GB showed limited stock on the day we checked, so expect its price to move first.
Watch these prices on Black Friday
Black Friday lands on November 27 this year. On November 6, none of the 128GB or RTX 5090 picks carried an Amazon discount. The cheap-looking offers you’ll see sit on configurations that are wrong for LLM work, so treat a big red percentage on a MacBook or a Flow Z13 as a reason to read the memory line twice.
- Binned M5 Max MacBook Pro, 36GB: $3,899, 11% off list, sold by Amazon
- Flow Z13 with 32GB: 15% off list, sold by Amazon
- New MacBook Pro 14 with M6, available from November 6: 32GB maximum
All three cap out around 32B and can’t load a 70B model or gpt-oss-120b, so none of them is a deal for this job. The binned M5 Max is the worst trap, because it carries the M5 Max name but gives you 36GB at 460 GB/s. If a 128GB Mac, the Flow Z13 or the SCAR 18 drops before November 27, the links above show the new price, and our Black Friday homelab deals list tracks the mini PCs and GPUs we cover.
Find the memory each model needs
File size is the floor. Add a few gigabytes for the KV cache, and more for long context. The sizes below are the actual Q4 GGUF files on Hugging Face, so they match what you download.
| Model | Quant | File size | Smallest laptop pool that runs it |
|---|---|---|---|
| Llama 3.1 8B | Q4_K_M | 4.9 GB | 8GB GPU |
| Qwen3-14B | Q4_K_M | 9.0 GB | 12GB GPU |
| gpt-oss-20b | MXFP4 | 12.1 GB | 16GB GPU |
| Qwen3-30B-A3B | Q4_K_M | 18.6 GB | 24GB GPU |
| Qwen3-32B | Q4_K_M | 19.8 GB | 24GB GPU, short context |
| Llama 3.3 70B | Q4_K_M | 42.5 GB | 64GB unified memory |
| gpt-oss-120b | MXFP4 | 63.4 GB | 128GB unified memory |
| Qwen3-235B-A22B | Q2_K | 85.7 GB | 128GB (on Linux, raise the GTT limit) |
The jump from 24GB to 64GB is where laptops split into two camps, because nothing with a discrete GeForce GPU sits in between. Decide which side of that line your models live on before you compare anything else. For the full sizing math, including KV cache per context length, see how much VRAM you need to run an LLM.
Read bandwidth, not TOPS
Token generation reads every active weight once per token, so decode speed is capped by memory bandwidth divided by model size. A 42.5GB 70B file on 614 GB/s tops out near 14 tokens per second. The same file on 256 GB/s tops out near 6, whatever the NPU TOPS figure on the box says.
Measured numbers follow that rule. These are llama-bench decode results for Llama 2 7B Q4_0, the standard yardstick in the llama.cpp CUDA and Apple Silicon scoreboards. The Strix Halo row is a Vulkan scoreboard entry on a desktop board.
| Hardware | Bandwidth | Decode, tokens/s |
|---|---|---|
| RTX 5090 desktop (reference) | 1,792 GB/s | 290.0 |
| RTX 5090 Laptop | 896 GB/s | 156.5 |
| M5 Max, 40-core GPU | 614 GB/s | 119.9 |
| M5 Pro, 20-core GPU | 307 GB/s | 66.3 |
| Ryzen AI Max+ 395 (Vulkan) | 256 GB/s | 51.8 |
The laptop RTX 5090 is a cut-down desktop 5080 die with half the bandwidth of the desktop card, and it decodes at about half the speed. It still beats every unified-memory laptop on any model that fits in 24GB.
MoE models change the math. gpt-oss-120b is 63GB on disk but activates about 5B parameters per token, and a Ryzen AI Max+ 395 decodes it at 50.2 tokens per second in a published ROCm run. That is why the 128GB machines earn their price: big MoE models run fast once they fit.
Plan for thermals and battery
Decode is mostly bandwidth-bound, so GPU power matters less than you’d expect, down to a point. On the CUDA scoreboard, an HP Omen Max 16 with an RTX 5090 at 140 to 160W decodes within 1 to 4% of a 175W run, while the same laptop in its 95W power mode drops about 28%. Prompt processing is compute-bound, so it falls with every watt the chassis can’t cool.
Battery is the real hit on NVIDIA laptops. StorageReview ran an RTX 5090 Razer Blade 16 plugged in and on battery, and Phi output fell from 163 to 99 tokens per second while Mistral fell from 126 to 67. That is a 40 to 47% loss. Apple says the M5 Max MacBook Pro delivers the same performance on battery as plugged in, which is the main reason the Mac wins for people who work away from a desk.
Chassis size changes what you get from the same chip. The thick 18-inch gaming laptops run the 5090 at its full 175W, but you’ll want the charger for anything heavy. The 14-inch MacBook Pro throttles more than the 16-inch under long GPU loads. The Flow Z13 is a tablet that runs Strix Halo at lower sustained power than a desktop box, so expect its prompt processing to trail the desktop numbers above even though decode lands close.
Choose between the four picks
MacBook Pro M5 Max with 128GB
This is the pick when the model is bigger than 64GB or you need it to run on battery.
The pick is the 14-inch with the 18-core CPU, 40-core GPU, 128GB and a 2TB SSD. A 14-inch 128GB M5 Max ran Llama 3.3 70B Q4_K_M at 9.95 tokens per second in LM Studio, per ifanr, which is about 70% of the bandwidth ceiling. gpt-oss-120b and Qwen3-235B at Q2 also load.
macOS caps how much memory the GPU can wire, and on a 128GB M5 Max llama.cpp reports about 107GB usable by default. A sysctl raises it by a few more gigabytes until the next reboot, and our local AI mini PC guide covers that setting for Macs and Strix Halo alike.
Two configurations catch buyers out. 128GB only comes with the full 40-core M5 Max, and the binned 32-core chip is capped at 36GB and 460 GB/s, per Apple’s spec page. That is the $3,899 MacBook Pro Amazon is discounting, and you should skip it for LLM work. If 64GB is enough, the 14-inch M5 Pro with 64GB costs $4,199 but has half the M5 Max bandwidth.
Buy the 16-inch if you run long agent jobs, because it holds clocks better. The 16-inch with 128GB and 2TB costs $7,699.99 from Adorama, which had limited stock on November 6. Skip the Mac entirely if you need Linux: Asahi Linux does not run on M5 Macs, so this is a macOS machine for llama.cpp, MLX, Ollama and LM Studio.
Expercom, the seller on this listing, had the 14-inch config in normal stock on November 6, while every 16-inch 128GB and 2TB listing still in stock was flagged as limited.
ASUS ROG Flow Z13 with 128GB
This is the cheapest way to fit gpt-oss-120b in a bag, and the one to buy if you live on Linux.
- Ryzen AI Max+ 395, Radeon 8060S with 40 CUs
- 128GB LPDDR5X-8000 soldered, 256 GB/s
- 13.4-inch 2-in-1 tablet, 1.2 kg, 70Wh battery, 200W adapter
Bandwidth is lower than the Mac’s, so a dense 70B Q4 model has a ceiling near 6 tokens per second. MoE models are where it pays off: gpt-oss-120b decoded at 50 tokens per second on the same chip in a Framework Desktop, and you should expect a little less in the tablet.
Linux is the other reason to buy it. On Windows, AMD caps GPU memory at 96GB through Variable Graphics Memory. On Linux the GPU uses GTT instead, and AMD’s ROCm tuning guide says to keep the BIOS carve-out small and raise the TTM page limit, which published runs put at a safe 108 to 116GB on a 128GB unit. ROCm wants kernel 6.18.4 or later, or the patched Ubuntu 24.04 HWE and OEM kernels, while Fedora 43 and Ubuntu 26.04 work as installed.
Skip it if you want a desk machine, because the 128GB is soldered and the tablet can’t sustain desktop Strix Halo power. The Ryzen AI Max+ 395 mini PCs are cheaper and cooler for that.
HIDevolution is a third-party reseller that builds this 128GB config, and it had limited stock on November 6, so check the count before you plan around it.
ASUS ROG Strix SCAR 18 with RTX 5090
This is the fastest laptop here for anything that fits in 24GB, and the only pick Amazon sells itself.
- RTX 5090 Laptop, 24GB GDDR7, 896 GB/s, 175W with Dynamic Boost
- Core Ultra 9 275HX, 32GB DDR5 in two SO-DIMM slots, 64GB max
- 380W adapter, 90Wh battery, 3.3 kg
The 5090 Laptop decodes Llama 2 7B at about 157 tokens per second. A Qwen3.8-27B Q4 build decoded at about 41 tokens per second on short prompts in a published test. With a 4-bit KV cache, a 256K context still fit in 24GB, and you’d still get about 36 tokens per second with 200K filled and speculative decoding on.
CUDA is the other reason to buy it. vLLM, SGLang, Unsloth fine-tuning and most research code assume an NVIDIA GPU, and the 18-inch chassis is rated for the full 175W, which keeps prompt processing fast. For the engine trade-offs, see Ollama vs vLLM vs llama.cpp.
Skip it if your models are 70B, because they won’t fit and spilling layers to DDR5 drops speed hard. If you want more system RAM for MoE offload, the Razer Blade 18 with an RTX 5090 and 64GB takes up to 128GB of DDR5 per Razer, but it costs $5,999.99 and showed limited stock.
Amazon ships and sells this one, which also makes it the easiest of the four to return if the fan noise under load bothers you.
MSI Vector 16 HX AI with RTX 5080
This is the value pick for models up to 14B and gpt-oss-20b.
- RTX 5080 Laptop, 16GB GDDR7, 896 GB/s, 175W with Dynamic Boost
- Core Ultra 9 275HX, 32GB DDR5, two slots, 96GB max
- 16-inch, 2.7 kg, 330W adapter
The RTX 5080 Laptop uses the same GB203 die and the same 256-bit bus as the 5090 Laptop. It loses 8GB of VRAM and some CUDA cores, not bandwidth, so models up to 14B and gpt-oss-20b decode at a similar speed to the SCAR 18 while prompt processing runs slower.
The 96GB RAM ceiling helps with MoE models. llama.cpp can keep the expert layers in system RAM and the rest on the GPU, which runs Qwen3-30B-A3B well, and our llama.cpp guide covers the offload flags.
On a tighter budget, the Vector 16 HX AI with an RTX 5070 Ti costs $1,599 from Newegg Business. It has 12GB at 672 GB/s, a Core Ultra 7 255HX and only 16GB of RAM, which is enough for 8B and 14B models and not much more. Skip both if you’ll want 32B models within a year.
Snow Bell sells this 32GB and 1TB configuration as a third-party seller, so read the listing’s warranty terms before you buy.
Know what RTX Spark laptops change
NVIDIA’s RTX Spark laptops became available from October 16 with up to 128GB of LPDDR5X at about 300 GB/s, per StorageReview, on a 45 to 80W chip. That is Strix Halo capacity with CUDA, and on Windows the MSI Prestige N16 decoded gpt-oss-120b at 65.4 tokens per second in a test Tom’s Guide reported.
We didn’t make one a pick for two reasons. The cheapest 128GB model in stock on Amazon, the ASUS ProArt P16 H7607BA, costs $8,099, more than the 128GB MacBook Pro with twice the bandwidth. And on native Linux the same MSI laptop fell to 13.9 tokens per second plugged in, which the testers traced to a Windows-only NVIDIA power controller. WSL ran at full speed. Our RTX Spark Linux explainer tracks the driver side.
HP’s ZBook Ultra G3a with the Ryzen AI Max+ PRO 495 and 192GB was not listed on Amazon US on November 6. Its bandwidth is only about 7% higher than the 395, see Ryzen AI Max 495 vs 395, so it won’t change the order above.
Check Linux support first
| Platform | Linux status | What to know |
|---|---|---|
| Apple M5 Max | No | Asahi Linux supports M3 since September 2026, without GPU acceleration; M4 and M5 are still in development |
| AMD Strix Halo | Yes | ROCm wants kernel 6.18.4+ or patched Ubuntu 24.04 kernels; Vulkan works in llama.cpp too |
| NVIDIA RTX 50 laptops | Yes | NVIDIA driver works; hybrid graphics uses PRIME offload, or set the MUX switch to dGPU only |
| NVIDIA RTX Spark | Partly | Runs, but one tested laptop decoded about 79% slower plugged in on native Linux; WSL matched Windows |
If Linux on bare metal is a hard requirement, the choice narrows to the Flow Z13 and the two GeForce laptops. Plan on the NVIDIA proprietary driver for the RTX 50 machines, and on a current kernel for the Z13.
Buy for the largest model you plan to run
Up to 14B, the RTX 5080 Vector is the value pick. Up to 32B, or if you need CUDA for training, buy the RTX 5090 SCAR 18 and keep it plugged in. For gpt-oss-120b on Linux, the Flow Z13 is the cheapest way in, and for 70B dense models or real work on battery, only the 128GB MacBook Pro holds up. If none of this needs to be portable, a desktop GPU or a Mac Studio gets you more memory per dollar.
