RTX Laptop GPU VRAM for Local AI: Every Card Listed

Short answer: an "RTX 4070 Laptop GPU" has 8GB of VRAM, not the desktop 4070's 12GB, and an "RTX 4090 Laptop GPU" has 16GB, not 24GB. The mobile part shares the marketing name and almost nothing else. Across the whole current line, the ceiling is 24GB (RTX 5090 Laptop) and the most common configuration on shelves is 8GB.

That single fact decides whether a $1,500-$2,800 machine can run the models you bought it for, and you cannot fix it later — laptop VRAM is soldered to the board. Below is every RTX 30, 40 and 50 series laptop GPU with its real memory capacity, taken from NVIDIA's own comparison sheet, plus what each tier will actually hold.

---

The Short Version

Three numbers to carry into a store: 8GB is the volume tier, 16GB is what an "RTX 4090 Laptop" or "RTX 5080 Laptop" gives you, and 24GB on the RTX 5090 Laptop is the hard ceiling for mobile GeForce.

• Mobile GPUs use narrower memory buses than their desktop namesakes. Narrower bus, fewer memory chips, less VRAM.
• The 5070 Laptop exists in both 8GB and 12GB versions of the same GPU. Two laptops, same sticker, different capability.
• Wattage varies by chassis on identical silicon. That changes speed, not which models fit.
• Nothing here is upgradeable after purchase.

---

RTX 50 Series Laptop VRAM

Every RTX 50 laptop GPU, as published by NVIDIA on its GeForce laptop comparison page:

VRAM, bus width and power ranges are from NVIDIA's laptop comparison page; bandwidth figures cross-checked against the Wikipedia RTX 50 mobile specification table, which agrees with NVIDIA on every capacity.

On the 5070's two capacities. The RTX 5070 Laptop launched at 8GB. Tom's Hardware reported that NVIDIA had quietly added a 12GB configuration of the same GPU — same 128-bit bus, using 3GB GDDR7 modules instead of 2GB ones. NVIDIA's comparison page now lists both. There is no marketing distinction, so the only reliable check is the individual laptop's spec sheet. If you search this and find pages confidently asserting 8GB, 12GB and 16GB for the 5070 Laptop, the first two are right for different SKUs and the third is someone reading a desktop 5070 Ti table.

For what fits in each of these tiers, our 8GB Ollama model picks and 12GB picks are the practical follow-up.

---

RTX 40 Series Laptop VRAM

The 40-series mobile line tops out at 16GB, and the 4070 Laptop is an 8GB card:

Capacities and bus widths from NVIDIA's laptop comparison page; bandwidth and power ranges from the Wikipedia RTX 40 mobile table, which matches NVIDIA capacity-for-capacity.

Two things worth staring at. First, the 4070 and 4060 Laptop are the same 8GB / 128-bit / 256 GB/s memory system — for inference, which is overwhelmingly memory-bound, the gap between them is much smaller than the badge suggests. Second, the 4090 Laptop's 16GB is less than a used desktop RTX 3090's 24GB, a card that launched in 2020. That is the whole argument against buying a flagship laptop for capacity, in one line.

---

RTX 30 Series Laptop (the Used Market)

If you are buying second-hand, the only 30-series laptop parts worth considering for local AI are the 16GB ones:

Source: Wikipedia's GeForce RTX 30 mobile specification table. The RTX 3080 Laptop is the trap here — it shipped in both 8GB and 16GB versions, and used listings frequently do not say which. A 16GB 3080 Laptop is a genuinely reasonable cheap local-AI machine; an 8GB one is a 3070 with a better sticker.

---

Mobile vs Desktop, Side by Side

Every mobile part in the current generation has equal or less VRAM than the desktop card with the same name — never more.

Desktop capacities from the Wikipedia RTX 40 and RTX 50 desktop specification tables; laptop capacities from NVIDIA as above.

The bandwidth gap is wider than the capacity gap. A desktop RTX 5090 moves 1,792 GB/s; the 5090 Laptop moves 896 GB/s — exactly half. A desktop 5070 moves 672 GB/s against the 5070 Laptop's 384 GB/s. Since single-stream token generation is bound by how fast the GPU can stream weights out of memory, that ratio is roughly the ratio of tokens per second you should expect for the same model.

This is why our per-card guides — for example the RTX 4070 model picks — are explicitly about desktop silicon. Do not transfer those numbers to a laptop with the same name.

---

The TGP Problem: Same Chip, Different Laptop

Two laptops with an identical RTX 4070 Laptop GPU can differ by 80W of power budget — NVIDIA publishes the part at 35-115W of GPU subsystem power — and they will not perform the same on a long generation.

What that does and does not change:

• It does not change what fits. 8GB is 8GB in a 35W ultrabook and a 115W desktop replacement. Model selection is identical.
• It does change sustained throughput. Short bursts look similar because the chassis has thermal headroom banked. A five-minute generation is where a thin laptop settles to a lower clock and stays there.
• It is a per-laptop specification, not a per-GPU one. Manufacturers publish it as "Maximum Graphics Power" or "Total Graphics Power" in the detailed spec sheet, often several clicks deep. If a listing will not tell you the wattage, assume the bottom of the range.

We are not publishing measured tokens per second per TGP band, because we do not have a bench of otherwise-identical laptops at 60W and 140W to measure honestly. Anyone who publishes a clean per-wattage table without describing that bench is guessing. What we can give you instead is the arithmetic ceiling and a way to measure your own machine.

The bandwidth ceiling. For batch-1 decoding, the GPU must read every active weight from VRAM once per token, so:

max tokens/sec ≈ memory bandwidth (GB/s) ÷ model size in VRAM (GB)

This is a first-order upper bound, not a prediction. Real throughput lands below it — attention cache reads, kernel overhead and thermal throttling all take their cut — and it is only valid while the whole model is in VRAM. It is still the single most useful number when comparing laptops, because it is set by the memory system rather than by the review sample's cooling.

Model sizes are the Q4_K_M figures published on each model's page in the Ollama library; ceilings are computed from the bandwidth column, not measured. Expect real numbers meaningfully below these, and treat "does not fit" as "will spill to system RAM and collapse to single-digit tokens per second."

To measure your own laptop rather than trusting anyone's table, run ollama run --verbose llama3.1:8b — the --verbose flag is documented as "Show timings for response" and prints the eval rate after each generation. Do it on a long prompt, twice, several minutes apart. The second number is the one that matters.

---

What Each VRAM Tier Actually Runs

8GB runs 7-8B models comfortably and nothing above 14B; 16GB is the first tier that feels unconstrained; 24GB is the mobile ceiling and still cannot hold a 70B model.

Leave headroom. The number above is the model file, and the KV cache grows with context length on top of it — a 32B model that exactly fills 24GB at 2K context will not survive a 32K one. Our VRAM calculator does that arithmetic per model and context length, and the Ollama model RAM/VRAM table lists footprints per tag.

If you need more than 24GB in something you can carry, GeForce mobile has no answer and unified memory is the only route — see the Apple Silicon AI buying guide for what 64GB and 128GB configurations cost and deliver.

---

How to Verify Before You Buy

Three checks, in order, on the exact SKU — not the GPU family.

1. Read the VRAM off the laptop's own spec sheet, not the GPU name. If the listing says "RTX 5070 Laptop" without a capacity, that is a listing you cannot buy from safely.
2. Find the Maximum Graphics Power figure. It sits in the detailed specifications on manufacturer sites. Two laptops with the same GPU and a 40W difference are different products.
3. On a machine you already own, confirm what you actually have: nvidia-smi --query-gpu=name,memory.total --format=csv prints the real device name and total VRAM in one line, on Windows and Linux alike.

If the answer comes back below what you need, do not fight it with quantization alone. Compare against a desktop before committing — the used GPU buying guide covers what a 24GB card costs second-hand, and GPU prices and the memory shortage explains why every VRAM tier got more expensive in 2026.

---

Verdict

Buy a laptop for local AI because you need it to be portable, never because you need capacity.

1. Check the number, not the name. The 4070 Laptop is 8GB, the 4090 Laptop is 16GB, and the 5070 Laptop is whichever of 8GB or 12GB the specific machine shipped with.
2. 8GB is the volume tier and it is a real constraint. It runs 7-8B models well and stops there.
3. 16GB is the first comfortable tier, which in practice means an RTX 4090, RTX 5080 or 16GB RTX 3080 Ti laptop.
4. 24GB is the ceiling, and it is still less memory than a six-year-old desktop card you can buy used for a fraction of a 5090 laptop.
5. Wattage decides speed, VRAM decides possibility. Get the capacity right first; you can live with a slower machine, not with a model that will not load.

Whether the laptop is the right purchase at all is a separate argument, and we made it at length in the honest laptop guide for local AI. This page exists so that if you do buy one, you know exactly what you are getting before the return window closes.

---

Sources & Method

Every capacity in this article was taken from a primary or reference source, not from retail listings:

• NVIDIA GeForce laptop comparison — first-party "Standard Memory Configuration", memory interface width, and GPU subsystem power ranges for the RTX 40 and RTX 50 mobile lines. This is the authority for capacity.
• Wikipedia GeForce RTX 30 / 40 / 50 series specification tables — memory bandwidth, memory clocks and desktop comparison figures. Checked against NVIDIA on every overlapping value; they agree.
• Tom's Hardware: "Nvidia quietly launches 12GB RTX 5070 laptop GPU" — the 12GB RTX 5070 Laptop launch, which is why that row has two capacities. NVIDIA's own comparison page lists both capacities for that part, so the two-SKU situation is confirmed first-party.
• Ollama model library — Q4_K_M file sizes for each model tag quoted.

What we did not do: we did not bench laptops. There are no measured tokens-per-second figures on this page, and the ceiling table is arithmetic derived from published bandwidth. When we have laptops on the bench at known wattages, we will publish measurements and label them as such.

---

FAQ