Best Local LLMs for Laptops (by RAM)

Which local LLM actually runs on your laptop?

The honest answer depends on one number: how much memory your machine has. Below is the same tier logic from the Private AI Blueprint — matched to specific open models, with what each tier is genuinely good at and where it falls short.

No affiliate links. No pretending a laptop-sized model beats a frontier one. Model names belong to their respective projects.

Read this before you choose

Find your RAM figure first

Every recommendation below hangs off one number. On Windows: Task Manager → Performance → Memory. On a Mac: Apple menu → About This Mac. Write it down before you download anything.

Assume quantised versions

The sizes above assume 4-bit quantised builds — the standard format for running models locally. A full-precision copy of the same model needs roughly four times the memory.

Leave headroom

A model wants most, not all, of your memory. Rough rule: keep about 4 GB free for the operating system and whatever else you have open.

Graphics memory changes the maths

A dedicated GPU or Apple Silicon shared memory moves you up a tier in practice. A CPU-only laptop can still run Tier 1 and Tier 2 comfortably.

The four tiers, and what to run in each

Find your RAM figure, read that row, ignore the rest.

Tier 1 — Tight but workable

8 GB RAM

Llama 3.2 3BPhi-3 Mini 3.8BQwen2.5 3B

Format: 4-bit quantised (Q4)
Speed: Fast enough to read along with
Good for: Drafting, rewriting, summarising short documents, tidying notes. Keep other heavy apps closed.
Worth knowing: Long documents and multi-step reasoning are where these smaller models start to slip.

Tier 2 — The comfortable spot

16 GB RAM

Llama 3.1 8BQwen2.5 7BMistral 7BGemma 2 9B

Format: 4-bit quantised (Q4)
Speed: Comfortable conversational pace
Good for: Everyday assistant work: document Q&A over your own files, longer drafting, structured summaries, day-to-day research.
Worth knowing: This is the tier most people should aim for before buying anything new.

Tier 3 — Room to be picky

32 GB RAM

Qwen2.5 14BGemma 2 27B (Q4)Mistral Small

Format: 4-bit or 5-bit quantised
Speed: Slower per word, noticeably better answers
Good for: Careful analysis, longer context, code and technical reading, running a chat model and a document index side by side.
Worth knowing: Bigger is not automatically better — check whether the extra wait earns its keep.

Tier 4 — Serious local work

64 GB RAM or more

Llama 3.3 70B (Q4)Qwen2.5 32BMixtral 8x7B

Format: 4-bit quantised, GPU or Apple Silicon strongly preferred
Speed: Deliberate — think 'send it a task and come back'
Good for: The closest a laptop gets to hosted-assistant quality, with everything still on your own disk.
Worth knowing: Memory bandwidth matters as much as capacity here. On a CPU-only machine this tier is patience-testing.

Picking the model is the easy part.

The rest is installing it, pointing it at your own documents, and knowing exactly what does and doesn't leave your machine. That's the Private AI Blueprint: a 32-page guide, 8 printable one-pagers, and three ready-made assistants — about an hour, start to finish.