Published 2026-09-27 · Last updated 2026-09-27

Can this laptop run a local LLM?

The short answer

A laptop can run a local LLM if its RAM fits the model: 8B models at Q4 need about 6GB, 13B about 10GB, 32B about 20GB, and 70B about 40GB — plus 2GB of OS overhead. Any 2026 laptop with 16GB of RAM runs 8B–13B models comfortably; 70B models need 64GB machines like a MacBook Pro M4 Max or a Strix Halo laptop.

Key numbers

  • RAM needed for local LLMs (Q4, 8K context): 8B ≈ 6GB, 13B ≈ 10GB, 32B ≈ 20GB, 70B ≈ 40GB.
  • Generation speed is memory-bandwidth-bound: ~100 GB/s gives 8–15 tok/s on 8B Q4; 270+ GB/s gives 30+ tok/s.
  • 16GB of RAM is the practical minimum for local LLMs in 2026; 32GB is the comfortable tier.

RAM, not the NPU, decides whether a local LLM runs at all. The model's weights must fit in memory: multiply the parameter count by the quantization size (Q4 ≈ 0.6 bytes per parameter), then add the KV cache for your context length and roughly 2GB for the OS and runtime.

Speed is a separate question, and it is decided by memory bandwidth, not TOPS. A laptop with 100 GB/s bandwidth generates roughly 8–15 tokens/sec on an 8B Q4 model; a 270 GB/s MacBook Pro or Strix Halo machine hits 30+ tokens/sec on the same model.

The fastest way to get an exact answer for your machine: use our Local LLM & VRAM calculator — pick the model size, quantization, and context window and it returns the exact RAM requirement plus laptops that meet it.

Related questions

Can 16GB RAM run a local LLM?

Yes — up to 13B models at Q4 with a usable context window. 32B models need 24GB+; 70B needs 64GB.

Does the NPU run the LLM?

Usually not today — most local inference (llama.cpp, MLX) runs on CPU/GPU and is limited by memory bandwidth. The NPU matters for Copilot+ features and efficiency.

What quantization should I use?

Q4 is the best quality-per-GB tradeoff for laptops. Q8 roughly doubles memory use for a small quality gain; Q3 saves memory but visibly degrades answers.

Go deeper

Verified against our benchmark dataset and testing methodology. Facts as of 2026-09-27.