Problems · Before you buy
What laptop specs do I need to run AI models locally?
Local AI inverts the usual spec logic: memory decides what is possible and speed only decides how pleasant. A model either fits in your VRAM or unified memory, or it does not run acceptably at all, so the first spec is the memory ceiling and the second is memory bandwidth. Cores and clocks come third.
Three tiers as floors. Entry (small models, 7 to 8B quantised): 16GB system RAM with 8GB VRAM, or 16GB unified memory. Balanced (13 to 14B comfortably, 30B quantised at a stretch): 12 to 16GB VRAM or 32GB fast unified memory. No-compromise (30B plus, image and video generation): 24GB plus VRAM or 64GB plus high-bandwidth unified memory, with cooling that can hold a half-hour load.
First, notice when it happens
The causes, ranked
1. The memory ceiling: VRAM or unified, it decides everything
most commonHow to confirm it
Model size at your quantisation must fit inside GPU-accessible memory with room for context. A 7B model quantised needs roughly 5 to 6GB; 13B roughly 9 to 10GB; 30B-class over 20GB; and context windows add gigabytes on top.
The fix
Buy the memory tier for the largest model you honestly intend to run, plus context headroom. On unified-memory machines the system RAM IS the ceiling, which is why 32GB and 64GB configurations exist as AI specs, not luxury.
2. Memory bandwidth: the tokens-per-second dial
most commonHow to confirm it
Inference speed tracks memory bandwidth more closely than core count: the same model runs visibly faster on wider, faster memory.
The fix
Discrete GPUs bring their own fast VRAM. On unified machines, compare bandwidth between configurations rather than assuming RAM size implies speed; the difference between tiers is often the bandwidth, and it shows every generated token.
3. Sustained thermals: inference is a marathon in bursts
commonHow to confirm it
Long chats, agent loops and batch generation hold high load for minutes at a stretch. The settle point governs the second half of every long session.
The fix
For occasional prompts, any tier serves. For daily heavy use, demand the same sustained evidence you would for video exports: a fan, and a measured curve that holds. Fanless machines run models; they just run them slower after minute five.
4. Software path: match the machine to the ecosystem
commonHow to confirm it
NVIDIA's CUDA remains the widest tool path; Apple's unified memory runs surprisingly large models via its own stack; other paths improve monthly but verify your specific tools first.
The fix
Pick the model runner and tools you actually want, then buy the hardware that path supports today, not the one promised a roadmap. A day of research here outvalues a tier of hardware.
5. Storage: models are tens of gigabytes each
occasionalHow to confirm it
A modest local library of models, checkpoints and datasets consumes hundreds of gigabytes before work begins.
The fix
1TB is the practical floor for anyone collecting models; 2TB for image and video generation practice. NVMe speed also shortens the model-load wait you will feel daily.
6. The cloud alternative, priced honestly
occasionalHow to confirm it
If your AI use is bursty and online is acceptable, API pricing may beat a hardware tier for years of usage.
The fix
Local wins on privacy, offline work, iteration speed and fixed cost at heavy use. Do the arithmetic for your pattern before paying for the no-compromise tier; sometimes the right AI laptop is the balanced tier plus an API budget.
Measure it instead of guessing
The thermal half of an AI purchase is testable in the return window exactly like an export machine: run the deep sustained test and read the settle point, because that is the machine your tenth prompt of the evening talks to. The memory half is arithmetic you do before buying, and this article just did it with you.
What software cannot tell you here
- We cannot measure tokens per second for your model and runner; a quick benchmark with your own stack in the return window owns that number.
- VRAM size and bandwidth are read from the spec sheet and reviews, not from any browser test.
Asked, in people’s own words
- How much RAM do I need to run LLMs locally?
- Rule of thumb at common quantisation: the model's billions times roughly 0.7 gigabytes, plus a few for context and the system. So 7B fits in 8GB of GPU-accessible memory, 13B wants 10 plus, 30B-class wants 20 plus. Unified-memory machines count their system RAM toward this; discrete-GPU machines count VRAM only.
- Can I run AI models without a dedicated GPU?
- Yes, on CPU or integrated graphics, at reading speed for small models: workable for experimentation, slow for daily work. Unified-memory machines blur the line usefully, because their GPU shares the big system memory. For serious local use, the memory ceiling and bandwidth still rule.
- Is a gaming laptop good for AI work?
- Often the best value path: gaming machines carry exactly the discrete VRAM and cooling that inference wants. Check VRAM size first (8GB entry, 12 to 16GB balanced), then the sustained thermals, since AI sessions hold load longer than most game scenes.