Memory decides which local model a laptop can run. Speed comes second. A laptop GPU tops out at 24GB of VRAM, which stops 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. Pick the memory pool first, then the chassis that keeps it fed.
This guide sizes four laptops against real model files and measured llama.cpp numbers. It also covers what battery power and a thin chassis cost you. By October 2026, no GeForce laptop ships with more than 24GB of VRAM. Every spec below was verified against the maker spec pages.
Quick picks
| Laptop | Memory for models | Bandwidth | Price on Amazon | Best for |
|---|---|---|---|---|
| MacBook Pro 16 M5 Max, 128GB | about 107GB by default | 614 GB/s | $6,999.00 | 70B dense models, 120B MoE, battery work |
| ASUS ROG Flow Z13, 128GB | 96GB on Windows, about 108GB on Linux | 256 GB/s | $3,499.00 | gpt-oss-120b and 70B on a budget, Linux |
| ASUS ROG Strix SCAR 18, RTX 5090 | 24GB GDDR7 | 896 GB/s | $4,499.99 | fastest CUDA for models up to 32B |
| MSI Vector 16 HX AI, RTX 5080 | 16GB GDDR7 | 896 GB/s | $2,820.88 | 14B models and gpt-oss-20b at full speed |
Prices are the Amazon buy box on October 18, 2026. Only the SCAR 18 is sold by Amazon itself. The other three come from third-party sellers, noted in each pick.
How much memory each model needs
File size is the floor. Add a few gigabytes for the KV cache, 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. Nothing with a discrete GPU sits in between. For the full sizing math, including KV cache per context length, see how much VRAM you need to run an LLM.
VRAM or unified memory: bandwidth sets the speed
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.
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 from 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. It has half the bandwidth of the desktop card and 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. A Ryzen AI Max+ 395 decodes it at 50.2 tokens per second in a published ROCm run. That is why the 128GB machines are worth it: big MoE models run fast once they fit.
Thermals and battery: what a laptop costs you
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. The same laptop in its 95W power mode drops about 28%. Prompt processing is a different story. It is compute-bound and drops with every watt the chassis cannot cool.
Battery is the real hit on NVIDIA laptops. StorageReview ran an RTX 5090 Razer Blade 16 plugged in and on battery. Phi output fell from 163 to 99 tokens per second, and Mistral 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.
- Thick 18-inch gaming laptops are rated for the 5090 at its full 175W. Plan on the charger for anything heavy.
- The 14-inch MacBook Pro throttles more than the 16-inch under sustained GPU load. For long agent runs, buy the 16.
- The Flow Z13 is a tablet. It runs Strix Halo at lower sustained power than a desktop box, so expect slower prompt processing than the desktop numbers above. Decode should land close, since the memory bus is the same.
The four laptops
1. MacBook Pro 16 M5 Max with 128GB: the biggest models on battery
Buy this one when the model has to fit in memory and the work happens away from a wall socket.
- Chip: M5 Max, 18-core CPU, 40-core GPU
- Memory: 128GB unified, 614 GB/s
- Storage: 2TB SSD
- Price: $6,999.00 in Silver, sold by Adorama
This is the only laptop that runs a 70B dense model at a usable speed. 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. That is about 70% of the bandwidth ceiling. gpt-oss-120b and Qwen3-235B at Q2 also load.
Two configs catch people out. 128GB only comes with the full 40-core M5 Max. The binned 32-core M5 Max, like this $3,499 listing, is capped at 36GB and 460 GB/s, per the Apple spec page. Skip it for LLM work. If 64GB is enough, the 16-inch M5 Pro with 64GB costs about $4,599 but has half the bandwidth.
macOS caps how much memory the GPU can wire. On a 128GB M5 Max, llama.cpp reports about 107GB usable by default, roughly 84%. The sysctl below adds about 5GB more, and resets on reboot.
sudo sysctl iogpu.wired_limit_mb=114688The catch is Linux. Asahi Linux does not run on M5 Macs yet, so this is a macOS machine for llama.cpp, MLX, Ollama and LM Studio.
2. ASUS ROG Flow Z13 with 128GB: 70B and gpt-oss-120b for half the price
This is the cheapest 128GB machine on the list, and unlike the MacBook it runs Linux.
- Chip: Ryzen AI Max+ 395, Radeon 8060S with 40 CUs
- Memory: 128GB LPDDR5X-8000, soldered, 256 GB/s
- Build: 13.4-inch 2-in-1 tablet, 1.2 kg, 70Wh battery
- Price: $3,499.00, sold by HIDevolution, a third-party reseller
Same capacity class as the Mac for half the money. Bandwidth is lower, so a dense 70B Q4 model has a ceiling near 6 tokens per second. MoE models are where it shines: gpt-oss-120b decoded at 50 tokens per second on the same chip in a Framework Desktop. Expect a little less in the tablet.
It is also the best Linux pick here. On Windows, AMD caps GPU memory at 96GB through Variable Graphics Memory. On Linux the GPU uses GTT instead, which defaults to about half of RAM. The AMD ROCm tuning guide says to keep the BIOS carve-out small and raise the TTM page limit. The value is in 4KB pages, so 108GB is 108 times 262,144. On Fedora, set it as a kernel argument with grubby.
sudo grubby --update-kernel=ALL --args="ttm.pages_limit=28311552"On Ubuntu, append the same argument to GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, run sudo update-grub, then reboot. Published runs put the safe ceiling at about 108 to 116GB on a 128GB unit. Go higher and the OS runs out of memory. For ROCm, the AMD guide wants kernel 6.18.4 or later, or the patched Ubuntu 24.04 HWE and OEM kernels. Fedora 43 and Ubuntu 26.04 work out of the box.
The tablet form factor has a cost. The 128GB is soldered, and the chassis cannot sustain desktop Strix Halo power. For a desk-bound setup, the Ryzen AI Max+ 395 mini PCs are cheaper and cooler.
3. ASUS ROG Strix SCAR 18 with RTX 5090: the fastest CUDA laptop
Pick this one for raw decode speed on models that fit in 24GB.
- GPU: RTX 5090 Laptop, 24GB GDDR7, 896 GB/s, 175W with Dynamic Boost
- CPU and RAM: Core Ultra 9 275HX, 32GB DDR5 in two SO-DIMM slots, 64GB max
- Power: 380W adapter, 90Wh battery
- Price: $4,499.99, sold by Amazon
Anything that fits in 24GB runs faster here than on any Mac. 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 window still fit in 24GB, and decode ran at about 36 tokens per second with 200K filled, even with speculative decoding on.
CUDA is the other reason to buy it. vLLM, SGLang, Unsloth fine-tuning and most research code assume an NVIDIA GPU. The 18-inch chassis is rated for the full 175W, which keeps prompt processing fast.
The catch is the 24GB wall. A 70B model will not fit, and spilling layers to DDR5 drops speed hard. The two SO-DIMM slots only take 64GB. For the engine trade-offs, see Ollama vs vLLM vs llama.cpp.
4. MSI Vector 16 HX AI with RTX 5080: 16GB at full bandwidth
This is the value pick for models up to 14B.
- GPU: RTX 5080 Laptop, 16GB GDDR7, 896 GB/s, 175W with Dynamic Boost
- CPU and RAM: Core Ultra 9 275HX, 32GB DDR5, two slots, 96GB max
- Build: 16-inch, 2.7 kg, 330W adapter
- Price: $2,820.88, sold by a third-party seller
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. Models up to 14B and gpt-oss-20b decode at a similar speed to the SCAR 18. Prompt processing is 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. Our llama.cpp guide covers the offload flags.
On a tighter budget, the Vector 16 HX AI with an RTX 5070 Ti costs $1,749.99. It has 12GB at 672 GB/s and ships with 16GB of RAM. That is enough for 8B and 14B models, not much more.
Worth waiting for
Two platforms target the 128GB-plus tier this fall. Neither is on sale yet.
- NVIDIA RTX Spark laptops. Up to 128GB on a 256-bit bus with a Blackwell GPU. NVIDIA quotes about 300 GB/s. The Microsoft event on October 7 is expected to bring prices. They run Windows on Arm only. Our RTX Spark Linux explainer tracks the driver situation.
- HP ZBook Ultra G3a 16. Ryzen AI Max+ PRO 495 with 192GB, up to 160GB for the GPU. HP says October, price not announced. Bandwidth is only about 7% higher than the 395, see Ryzen AI Max 495 vs 395.
Intel Panther Lake laptops are not in the running. The Arc B390 iGPU decodes an 8B model at about 24 tokens per second, and the Arc B390 chips top out at 96GB of RAM.
Linux support by platform
| 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 | No | Windows on Arm only at launch |
Pick by 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. For 70B dense models, or real work on battery, only the 128GB MacBook Pro holds up. If none of these needs to be portable, a desktop GPU or a Strix Halo mini PC gets you more per dollar.
