Every laptop LLM question reduces to one number, and it is not the GPU’s name. It is how many gigabytes of video memory the chassis was sold with, because that is what you cannot change afterwards.
Quick answer: 8GB runs 7B and 8B models well and stops there. 12GB adds the 13-14B tier. 16GB reaches the mid-20s. 24GB — the RTX 5090 Laptop GPU — is the only mobile part that reaches 32B comfortably, and no laptop runs 70B without offloading to system RAM and waiting.
The 16GB desktop floor
NVIDIA GeForce RTX 4060 Ti 16GB
16GB GDDR6
16GB for far less than a 16GB laptop costs, in a machine where the memory is a part you can replace rather than a decision you live with.
Affiliate links. We may earn a commission at no extra cost to you. Amazon ships worldwide; Shopee SG covers Singapore and ASEAN.
Who this is for
You want models running on the machine you carry — a coding assistant on the train, a private summariser that never leaves the device, or a way to learn the stack without building a second computer. The constraint set is different from a desktop’s: you cannot add a card, you cannot add memory, and sustained power is rationed. What follows assumes all three.
The mobile tiers, from NVIDIA’s table
Manufacturer figures from NVIDIA’s laptop GPU comparison:
| Laptop GPU | Memory | Bus width | GPU subsystem power |
|---|---|---|---|
| RTX 5090 Laptop | 24 GB GDDR7 | 256-bit | 95–150 W |
| RTX 5080 Laptop | 16 GB GDDR7 | 256-bit | 80–150 W |
| RTX 5070 Ti Laptop | 12 GB GDDR7 | 192-bit | 60–115 W |
| RTX 5070 Laptop | 12 or 8 GB GDDR7 | 128-bit | 50–100 W |
| RTX 5060 Laptop | 8 GB GDDR7 | 128-bit | 45–100 W |
| RTX 5050 Laptop | 8 GB GDDR7 | 128-bit | 35–100 W |
Two traps live in that table. The RTX 5070 Laptop GPU ships in two memory configurations, so the model number alone does not tell you which tier you are buying — read the laptop’s own listing. And the wattage column is a range, not a value: the same GPU in a thin chassis and a thick one are two different products, and only the laptop maker says which you are getting.
Memory tier to model tier
Published download sizes at Q4_K_M, the quantization most people actually run:
| Model | Weights | Fits in |
|---|---|---|
| Phi-3 mini | 2.4 GB | 8 GB, easily |
| Mistral 7B | 4.4 GB | 8 GB |
| Llama 3.1 8B | 4.9 GB | 8 GB |
| CodeLlama 13B | 7.9 GB | 12 GB |
| Qwen 2.5 14B | 9.0 GB | 12 GB |
| Gemma 2 27B | 17 GB | 24 GB |
| Qwen 3 32B | 20 GB | 24 GB |
| Llama 3.3 70B | 43 GB | no laptop |
Weights are not the whole bill. The KV cache grows with context and is charged per token: around 0.125 MB per token on an 8B model and 0.188 on a 14B at FP16. A 14B at 32K context therefore wants roughly six gigabytes of cache on top of its nine of weights, which is how a model that “fits in 12GB” stops fitting the moment you paste in a long document. Add about 1.5GB of runtime overhead on top of both.
The same tiers on a desktop are covered in how much VRAM for local LLM, and the model-by-model detail is in the Ollama VRAM guide. That arithmetic is the single most useful thing to internalise before buying: the tier you need is set by your context length as much as your model size. Our VRAM calculator separates the two.
Check NVIDIA GeForce RTX 5070 Ti on AmazonBuy on Shopee SGWhere mobile silicon differs from the desktop part
The flagship name is not the flagship card. The desktop RTX 5090 carries 32GB; the laptop part of the same name carries 24GB. For LLM work that is the difference between 32B models with room to spare and 32B models with very little.
Narrow buses cost you tokens per second. Generation reads the whole model out of memory once per token, so throughput tracks memory bandwidth. The 128-bit parts at the bottom of the table are narrow by design, and that shows up directly as a lower ceiling on tok/s — see our methodology for how we model that relationship.
Apple’s portables answer this differently, since unified memory scales past any laptop GPU’s VRAM — our Mac versus NVIDIA comparison works through the trade.
Sustained power is rationed. A long generation is a long full-power draw. Laptops settle to a lower clock within the first minute and stay there; desktops with three to five times the power budget do not. Unplugged, almost every laptop cuts GPU power dramatically.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Which GPU should you buy, by tier?
- Coding assistant on 7-8B models, budget first? 8GB (RTX 5060 or 5050 Laptop). Genuinely good at this tier, and honest about its ceiling.
- Want the 13-14B tier, which is where quality jumps? 12GB — and confirm whether the RTX 5070 in that listing is the 12GB or the 8GB version.
- Long context, document work, room to grow? 16GB (RTX 5080 Laptop). Holds a 14B with a large cache, or a mid-20s model with a small one.
- The biggest models a laptop can hold? 24GB (RTX 5090 Laptop), which reaches 32B at Q4 and stops there.
- 70B, locally? Not on a laptop. Two desktop cards, as in the two-3090 guide, or rent.
Common mistakes
- Buying the 8GB RTX 5070 thinking it is the 12GB one. Same name, different part, and the difference is an entire model tier.
- Sizing for weights and forgetting context. A 14B fits 12GB at short context and does not at 32K. The cache is the part that surprises people.
- Reading desktop tok/s figures. Mobile parts have narrower buses and a fraction of the power budget. The numbers do not transfer.
- Assuming system RAM rescues you. Offloading layers to RAM works and is slow — slow enough to change what the machine is for.
- Choosing a thin chassis for a sustained workload. Generation holds the GPU at load for as long as it runs, which is precisely what thin cooling handles worst.
Verdict
| Target | Tier | Example part |
|---|---|---|
| 7-8B, coding and chat | 8 GB | RTX 5060 Laptop GPU |
| 13-14B, better reasoning | 12 GB | RTX 5070 Ti Laptop GPU |
| 14B with long context | 16 GB | RTX 5080 Laptop GPU |
| Up to 32B | 24 GB | RTX 5090 Laptop GPU |
| 70B | not portable | two desktop 24GB cards |
What the money buys on a desk
NVIDIA GeForce RTX 3090
24GB GDDR6X
24GB used — the same capacity as the mobile flagship, at a fraction of the price of the laptop carrying it, with none of the power ceiling.
Affiliate links. We may earn a commission at no extra cost to you. Amazon ships worldwide; Shopee SG covers Singapore and ASEAN.
On a laptop, VRAM is the only specification you buy once and keep forever. Pick the tier for the model you will want in two years, not the one you are testing this week.
Frequently asked questions
How much VRAM does a laptop need for local LLMs?
Eight gigabytes covers the 7-8B tier comfortably, which is enough for a capable coding and chat assistant. Twelve opens up the 13-14B models. Sixteen is where long context stops being a problem, and twenty-four is the ceiling any laptop reaches.
Can a laptop run a 70B model?
Not in any way you would enjoy. The weights alone are several times the memory of the largest mobile GPU, so most of the model lives in system RAM and generation slows to a crawl. That tier needs two desktop cards or a rented machine.
Is the laptop RTX 5090 as good as the desktop one?
No. NVIDIA’s own specifications give the mobile part less memory than the desktop card and a power budget a fraction of the size. It is the best laptop option and it is not the same component.
Does a Mac make more sense for portable LLM work?
Often, yes, because unified memory can be configured far beyond any laptop GPU’s VRAM and large models fit in one device. The trade is software: CUDA-only tooling does not run there. Our Mac versus NVIDIA comparison works through where each wins.
Will running models constantly damage the laptop?
Not damage, but it runs hot and loud, and sustained thermal cycling is harder on a portable than on a tower. Expect the fans to be audible for as long as generation lasts, and expect clocks to settle below peak quickly.
References
- NVIDIA — GeForce laptop GPU comparison — memory, bus width and GPU subsystem power for every RTX 50 Series laptop part.
- NVIDIA — GeForce RTX 5090 — the desktop specifications the mobile flagship is measured against.
- Ollama — llama3.1 tags — published download sizes for the 8B rows.
- Ollama — qwen2.5 tags — published download sizes for the 14B and 32B rows.
- Our VRAM calculator — weights, KV cache and overhead separated, at any context length.
- Our methodology — how tokens per second are modelled from memory bandwidth.